[{"id":"L0_N15647","k":"d","L":0,"N":15647,"t":"A layer-0 ' Park' token detector, blind to meaning","l":"This layer-0 neuron is a lexical detector for the token ' Park'. It fires just as strongly on a surname, a street or a national park, because it reads the token embedding directly and has no context to use. Nearby tokens such as park, parks, parking and 公园 get about half its…","st":"tested","i":2,"c":9,"m":[["surname Park",11.7],["core ' Park'",10.4],["lowercase park",7.0],["inflected",5.2],["-ark rhyme",3.5],["merged token",2.8],["Parker",2.6],["translation",0.7],["synonym control",0.2],["no-cue control",0.1]]},{"id":"L10_N7650","k":"d","L":10,"N":7650,"t":"The 'more and more' neuron: comparatives in an iterative frame","l":"L10/N7650 detects one grammatical shape: a comparative sitting right after an iterative particle (Russian все, Spanish cada vez, German immer, English ever) or right after a copy of itself, as in 'more and more'. It does not care about direction or meaning, only adjacency and…","st":"tested","i":5,"c":7,"m":[["X-er and X-er",23.2],["ever + comparative",13.6],["increasingly",10.9],["growing + noun",5.9],["mismatched pair",3.3],["'or' pair",2.4],["plain comparative",1.0],["no cue",0.5],["tempo + participle",-1.3],["tempo + participle",-1.3]]},{"id":"L11_N12072","k":"d","L":11,"N":12072,"t":"The moment it clicks, or the penny drops","l":"Fires on the word where understanding lands: it clicked, the penny dropped, it dawned on me, the reality sank in. It works in Japanese, Chinese and Russian, and mostly ignores clicked buttons and sinking ships. Plain “I realized” or “I figured it out” gets a small fraction as…","st":"tested","h":"../notes/L11_N12072/","i":7,"c":8,"m":[["the penny dropped",24.2],["the whole proof clicked",21.6],["it didn't click",18.5],["ピンときた",17.5],["news sank in",16.5],["clicked the button",3.0],["I figured it out",2.59],["ship sank in storm",0.049]]},{"id":"L11_N3324","k":"d","L":11,"N":3324,"t":"A 'concerned with' detector that recognizes disguised synonyms","l":"This neuron fires on the word that completes 'concerned with', usually ' with' itself. It fires hardest when a person, a pronoun, or nothing at all is the subject ('Concerned with topology'). The common topic-scoping use, 'This course is concerned with', gets roughly half that…","st":"tested","i":5,"c":8,"m":[["fragment",21.1],["reflexive verb",18.5],["worried",18.4],["pronoun subject",17.0],["human subject",16.4],["inanimate subject",11.8],["preoccupied",11.7],["anxious, no spin",9.3],["scope verb",3.3],["control adj",1.3]]},{"id":"L12_N12088","k":"d","L":12,"N":12088,"t":"Forty-two, when it’s the answer to everything: a Hitchhiker’s Guide neuron","l":"A neuron that knows The Hitchhiker’s Guide to the Galaxy: the title, Douglas Adams, Zaphod, and above all the 42 that ends “the answer to life, the universe, and everything.” A 42 on a bus gets only a small bump. It knows the book by its words, so an ordinary hitchhiker or the…","st":"tested","h":"../notes/L12_N12088/","i":8,"c":9,"m":[["answer to everything",68.5],["Douglas Adams",45.5],["The answer is 42",34.0],["roadside hitchhiker",29.5],["Jackie Robinson's 42",11.9],["bus 42",9.94],["The answer is 17",3.22],["Mr. Adams, humorist",0.445]]},{"id":"L12_N14823","k":"d","L":12,"N":14823,"t":"A neuron that favors textbook radixes like 'base three'","l":"This neuron fires on the word that names a number system's radix, as in 'base three', 'ternary', 'Basis 3' or 三进制. Its strength depends on which radix is named. The small radixes of textbook conversion exercises (three through seven) score highest, and binary, decimal and hex…","st":"tested","i":7,"c":7,"m":[["three (core)",18.9],["four",16.8],["26 (coding)",16.2],["seven",14.4],["thirteen",6.8],["sixteen (hex)",6.6],["two (binary)",4.9],["camp (control)",1.9],["ten (decimal)",-0.1]]},{"id":"L12_N3240","k":"d","L":12,"N":3240,"t":"Bengali surname string detector, firing negative on the surname token itself","l":"This neuron is really a negative detector for eastern Indian surnames. It plunges on the last token of a Bengali, Odia or Assamese surname spelled the Indian way, such as Maity, Ghosh or Chakraborty, even with no context at all. Place names, Bengali given names and other Indian…","st":"tested","i":5,"c":6,"m":[["Indian given + Western",4.2],["Western control",0.0],["Gujarati control",-0.3],["Bangladeshi Muslim",-1.0],["pan-Indian given + Ghosh",-2.6],["anglicized Bengali",-5.0],["Bihari neighbour",-7.8],["Ghosh bare",-11.1],["Chakraborty",-13.9],["Bengali given + Maity",-28.5]]},{"id":"L12_N6594","k":"d","L":12,"N":6594,"t":"'The number and' detector whose negative pole names zoonotic infections","l":"This neuron watches for a count-of-things noun such as 'number' or 'amount' followed straight away by 'and', a comma or 'or'. It fires on the coordinator itself, before the second attribute appears, and holds through 'type of'. Other coordinated attributes like 'size and shape'…","st":"tested","i":5,"c":8,"m":[["number (core)",14.9],["amount (core)",11.2],["quantity",7.9],["frequency (halo)",4.4],["size (halo)",4.1],["date (halo)",3.1],["already coordinated",2.5],["identifier sense",1.2],["name (control)",1.1],["zoonosis (control)",-47.0]]},{"id":"L13_N25183","k":"d","L":13,"N":25183,"t":"A MetaFilter template matcher that decides the byline's next token","l":"This neuron recognizes a handful of lowercase MetaFilter page-template strings and fires on the token that completes them: 'Filter' after 'Meta', ' Subscribe' after a clock time, ' inside' in '[more inside]', and the ' on' or ' -' that follows a byline time or date. It is…","st":"tested","i":3,"c":8,"m":[["posted by (core)",29.9],["Capital P",17.0],["Answer by",12.6],["Drupal Submitted by",11.6],["name only",11.6],["Written by",9.2],["nonsense verb",7.4],["no byline",4.9],["deadline prose",3.9],["prose 'see you'",3.5]]},{"id":"L13_N8283","k":"d","L":13,"N":8283,"t":"'One more on the pile' phrases, and wrapped code-comment indentation","l":"This neuron has two unrelated jobs. Its strong negative pole fires on fronted phrases that pile one more misfortune onto a list ('On top of that', 'To add insult to injury', 'Para piorar'). It builds token by token and reaches about −40 on the closing word. Invented versions…","st":"tested","i":7,"c":8,"m":[["nonsense control",0.2],["time control",-0.7],["plain escalator",-4.0],["correlative",-5.9],["additive connective",-7.4],["novel metaphor",-11.2],["invented idiom",-29.4],["stock idiom",-31.8],["as-if frame",-33.0],["stock idiom",-33.8]]},{"id":"L14_N11264","k":"d","L":14,"N":11264,"t":"A Christmas Day neuron: the 25 in December 25, minus Saint John","l":"This neuron fires hardest on the digit that completes a December 25 date, in prose, tables or code, and harder still when Christmas is named nearby. Carol titles, reindeer, mistletoe and the word Christmas fire it at about half strength. Its gate detects festival names of any…","st":"tested","i":9,"c":8,"m":[["nativity",37.8],["Christmas",36.5],["other holiday",36.5],["nonsense",33.8],["solstice",32.2],["Hanukkah",32.0],["Catalan Christmas",30.1],["no festival",28.1],["Saint John",26.9]]},{"id":"L14_N5420","k":"d","L":14,"N":5420,"t":"One grain of rice on the first square: a neuron for the chessboard legend","l":"Fires on the legend where the inventor of chess asks a king for one grain of rice on the first square, then double on each square after. It peaks on the props (grain, square, chessboard) before any doubling is said and never checks the arithmetic. Other doubling fables get half…","st":"tested","h":"../notes/L14_N5420/","i":8,"c":9,"m":[["lesson plan",81.0],["king's rice",74.0],["rajah's rice",47.2],["Spanish legend",43.2],["rice per cup",35.8],["doubling penny",32.0],["virus doubling",4.75],["chess club rice",2.3]]},{"id":"L14_N5865","k":"d","L":14,"N":5865,"t":"A place-name lookup: Po plain and Great Lakes negative, Kaqchikel positive","l":"This neuron fires on the last piece of a multi-token name, once the model knows which name it is reading, and largely ignores the surrounding sentence. Its strong side is negative and geographic. Towns of the central and lower Po plain push it down hardest, to about -27, and…","st":"tested","i":4,"c":6,"m":[["Kaqchikel",7.1],["positive name",4.7],["off-plain Italy",0.7],["nonsense",-0.1],["Wisconsin",-0.4],["Veneto edge",-3.2],["S. Michigan",-20.9],["Po plain",-21.0],["Po plain",-23.1],["Emilia",-24.5]]},{"id":"L14_N7631","k":"d","L":14,"N":7631,"t":"Everyone's hat but your own: a neuron for hat puzzles","l":"Fires for hat puzzles, where prisoners or gnomes each wear a colored hat, see everyone else's, and must work out their own. It works across languages, gives about half as much to muddy foreheads and blue-eyed islanders, and ignores other famous puzzles. A weaker second home…","st":"tested","h":"../notes/L14_N7631/","i":8,"c":7,"m":[["German hat puzzle",45.8],["gnomes, Russian",41.0],["see others' hats",34.2],["Bertrand duopoly",24.9],["mud on forehead",16.2],["blue-eyed islanders",12.8],["magician's hat",0.02],["knights and liars",-1.22]]},{"id":"L14_N7761","k":"d","L":14,"N":7761,"t":"“But, you may say…”: a neuron for the reader who talks back","l":"Fires when a writer hands the floor to an imagined critic: “But, you may say,” “It will be objected that,” “Now, I know what you're thinking.” It peaks on the speaking verb or the comma or “that” just after it, works across many languages (a Spanish “Pero, dirá usted,” reaches…","st":"tested","h":"../notes/L14_N7761/","i":7,"c":9,"m":[["Pero, dirá usted,",52.5],["But, you'll say",38.0],["Vous me direz",28.6],["you may object",27.1],["it will be said",22.9],["you may wonder",11.3],["defense will argue",2.55],["but you said",0.9],["in my mind",-5.06]]},{"id":"L15_N10410","k":"d","L":15,"N":10410,"t":"Butterfly flapping its wings, and events in one part of a whole","l":"This neuron fires on the image of a butterfly flapping its wings, peaking on the wings token in any language. A chaos or weather-forecast frame pushes it higher, to about +54. Other flapping creatures, and the bare phrase 'butterfly effect', give weaker responses. A second route…","st":"tested","i":6,"c":6,"m":[["butterfly wings",33.5],["one-part template",21.9],["eagle wings",17.5],["seagull wings",17.0],["butterfly, no wings",14.8],["named effect",11.9],["supply chain",11.4],["technical",8.1],["difference between",5.8],["control",0.3]]},{"id":"L15_N10957","k":"d","L":15,"N":10957,"t":"A neuron for safe rooms, storm shelters and blast doors","l":"This neuron responds to the last word of a compound that names a structure built to withstand force: the 'room' in 'safe room', the 'shelter' in 'storm shelter', and 'blast door', 'hurricane shutters' or 'bulletproof'. The compound does most of the work. A tornado or war setting…","st":"tested","i":4,"c":8,"m":[["tornado safe room",34.2],["safe room",27.9],["storm shelter",23.4],["blast door",22.5],["hurricane shutters",19.5],["fortified wall",14.7],["fire door",5.1],["bare room",0.6],["flood barrier",0.3],["animal shelter",-0.0]]},{"id":"L15_N4975","k":"d","L":15,"N":4975,"t":"A neuron for putting things off until the last moment","l":"This neuron fires on the endpoint of a phrase that stretches time to the latest possible point, such as 'until the last moment' or 'last minute', and on their equivalents in Italian, Portuguese and Chinese. The phrase alone carries most of the signal. A procrastination verb such…","st":"tested","i":6,"c":8,"m":[["until last moment",31.2],["till last minute",28.6],["last possible second",24.9],["eleventh hour",19.6],["at last moment",18.9],["final moment",15.3],["as long as",11.9],["until Friday",9.6],["until midnight",9.6],["nonsense control",9.6]]},{"id":"L16_N10735","k":"d","L":16,"N":10735,"t":"A detector for famous names of the Cinderella story","l":"This neuron fires when the model recognizes a well-known name for the Cinderella story. The strongest responses are to a handful of memorized variant titles: Donkeyskin, Aschenputtel, Yeh-Shen and Ye Xian, which reach about +55. Plain Cinderella, its translations and other…","st":"tested","i":6,"c":8,"m":[["Donkeyskin",47.2],["Yeh-Shen",44.8],["Aschenputtel",40.2],["translation",29.1],["plot motif",20.6],["Cinderella",19.6],["other tale",15.9],["near-synonym",8.2],["obscure variant",-0.3],["control",-0.7]]},{"id":"L16_N4186","k":"d","L":16,"N":4186,"t":"Memorized article-spinner lexicon, negative on dialects of known languages","l":"This neuron recognizes the thesaurus 'article spinner' dialect of essay mills, where laws become Torahs. It is a memorized lexicon, not a rule: attested spinner forms fire, the plain words they replace do not, and invented doubled-s words give nothing. Non-English strings fire…","st":"tested","i":9,"c":6,"m":[["ESL form",21.2],["spinner non-word",20.4],["spinner doubled-s",18.6],["fake doubled-s",13.0],["spinner doubled-s",12.7],["wrong-sense real word",12.4],["Scots",4.0],["rare but fitting",1.9],["plain control",1.8],["misspelling",1.5]]},{"id":"L17_N10171","k":"d","L":17,"N":10171,"t":"Echo chambers, named or spelled out as confirming beliefs","l":"This neuron tracks selective exposure: people seeing only information that fits what they already believe. It responds in two ways. The name itself ('echo chambers', 'filter bubbles') peaks on the concept word, near +26. A spelled-out paraphrase ('content that confirms their…","st":"tested","i":5,"c":7,"m":[["partisan name",28.9],["confirms beliefs",25.5],["echo chamber",25.0],["filter bubble",23.1],["challenges beliefs",19.6],["opinions",10.8],["narrowed belief",9.9],["hobbies",6.5],["no-frame control",2.3],["literal chamber",1.1]]},{"id":"L17_N5921","k":"d","L":17,"N":5921,"t":"A neuron for stock formulas that an inscribed object reads","l":"This neuron fires on the opening words of a stock inscription, such as 'Here lies', 'To the Memory of', 'Keep off' or 'Property of', when the text presents them as what a written object says. Two cues multiply: a named object doing the saying (stone and metal count most) and a…","st":"tested","i":7,"c":8,"m":[["wedding cake",26.0],["memorial plaque",24.2],["gravestone",24.0],["wooden sign",22.2],["product label",20.6],["book page",19.6],["email",19.0],["nonsense",18.9],["human speech",14.7]]},{"id":"L17_N652","k":"d","L":17,"N":652,"t":"The puppet on the stage of a baby-psychology experiment","l":"This neuron fires on the staged social partner in a psychology experiment: the puppet, experimenter, confederate or doll that a subject watches, helps, lies to or judges. It is strongest for puppets shown to infants, children or animals (about +28), weaker for adult subjects…","st":"tested","i":7,"c":7,"m":[["infant study",27.1],["child study",23.9],["adult study",18.4],["dog study",17.0],["social-psych adults",16.5],["market",4.8],["nonsense",4.6],["party",3.7],["history",3.4]]},{"id":"L18_N6230","k":"d","L":18,"N":6230,"t":"Fires on the head or face a rock seems to resemble","l":"This neuron fires on the word that names what something looks like, such as 'head' in 'the rock resembles a man's head'. It is strongest when the subject is a natural thing that happens to look like something, such as a landform, fruit, insect marking or nebula, and when the…","st":"tested","i":8,"c":5,"m":[["rock",48.5],["nebula",47.5],["granite tor",45.5],["cloud",39.5],["building",34.0],["nonsense",29.6],["fruit",25.5],["abstract",24.5],["statue",19.9],["human subject",17.6]]},{"id":"L19_N10517","k":"d","L":19,"N":10517,"t":"Using up a stock on purpose so nothing goes to waste","l":"This neuron fires when text sets out to consume a stock on purpose so nothing is wasted. The clearest trigger is 'up' in 'a great way to use up', which reaches about +28. A second path fires when leftovers, byproducts or invasive species are named as the input to something new.…","st":"tested","i":6,"c":6,"m":[["use up",33.0],["make use of",24.8],["every last bit",23.6],["use all of",21.2],["plain use",19.1],["eat up",17.6],["burn through",16.4],["finish off",14.8],["show off",12.1],["no-use control",8.1]]},{"id":"L19_N15514","k":"d","L":19,"N":15514,"t":"The 'not if, but when' neuron: marks the open question","l":"This neuron fires on the word that names which question about an event is actually open. Its best case is 'not a matter of if, but when'. The dismissed 'if' already fires, and the favoured 'when' peaks near +50, whatever the topic. Plain 'whether' clauses and inevitability talk…","st":"tested","i":6,"c":8,"m":[["full idiom",58.5],["split frame",55.5],["no 'if' half",50.5],["bare contrast",47.2],["whether variant",45.8],["inevitability slot",44.2],["real question",40.8],["plain 'when'",3.4],["nonsense control",1.0]]},{"id":"L19_N9026","k":"d","L":19,"N":9026,"t":"A detector for English written backwards, plus fronted echo verbs","l":"This neuron detects letters that read like English written backwards. Two cues drive it: backwards-English letter runs such as 'eht' and 'gni', and a capital stranded at the end of a word ('dlroW olleH'). It responds to letter patterns rather than meaning, so reversed nonsense…","st":"tested","i":7,"c":6,"m":[["reversed + capital",40.0],["reversed pseudowords",22.6],["anagram drill",17.0],["reversed lowercase",14.1],["Pig Latin",5.3],["reversed Spanish",3.1],["word order",1.1],["countdown",0.5],["consonant junk",0.5],["forward control",-0.1]]},{"id":"L19_N9151","k":"d","L":19,"N":9151,"t":"A neuron for heaven on earth and a little bit of Italy","l":"This neuron fires when text says something from elsewhere, such as heaven, hell, God, paradise, Mars or Italy, is present here. Two frames carry it: '<realm> on earth' and 'a little bit of <place> right here'. It peaks on the host word 'earth' or 'here', and highest (about 47)…","st":"tested","i":8,"c":8,"m":[["bit of Italy",32.5],["bit of heaven",29.8],["paradise",27.4],["bit of Mars",26.6],["heaven",25.0],["hell",19.6],["mundane control",19.1],["abstract halo",16.8],["nonsense",13.0],["generic 'bit of'",11.1]]},{"id":"L1_N1962","k":"d","L":1,"N":1962,"t":"A 颜/彦 token detector whose colour label is closed-gate leak","l":"This unit is a token detector, not a colour detector. Its gate opens only for characters in the 颜/彦 family, whether they mean face, colour or a personal name, and its output there is strongly negative. Attention copies a standalone 颜 or 顏 one token forward, so a following 色 dips…","st":"tested","i":3,"c":8,"m":[["色 leak",0.5],["谚 same component",-0.1],["严 homophone",-0.1],["颈 same radical",-0.1],["天 control",-0.1],["顏 traditional",-11.1],["颜 core",-12.7],["彦 name",-17.1],["颜色 fused",-21.4],["彥 byte-split",-24.6]]},{"id":"L1_N20535","k":"d","L":1,"N":20535,"t":"Completes swing, sting, fling and cling from onset plus '-ing'","l":"This early-layer neuron completes a family of short words built from a consonant cluster plus '-ing': swing, sting, sling, fling, cling and wing. It fires on the rhyme token, and fires hardest (up to about 25) when the onset sat in the previous token. Strong w/l/St onsets drive…","st":"tested","i":3,"c":10,"m":[["split SW|ING",25.1],["split Sw|ings",21.4],["split St|ings",20.8],["whole word",15.6],["nonce w-onset",13.6],["r-cluster",10.3],["-ung vowel",9.2],["weak onset",2.3],["-ink control",0.5],["vowel-stem suffix",-0.9]]},{"id":"L1_N20736","k":"d","L":1,"N":20736,"t":"A Russia-word detector that fires negative, with a gate-off positive shadow","l":"This neuron detects the word Russia, not the idea of Russia. Whenever a token completes Russia or Russian in any script, English, Cyrillic, Chinese, Arabic or a code namespace like package ru, it swings sharply negative, reaching about -31 in code headers. Moscow, Kremlin and…","st":"tested","i":4,"c":9,"m":[["Romanian (halo)",0.7],["Prussian (relative)",0.6],["nonsense",0.6],["Ukrainian (halo)",0.5],["Russ (stem)",0.4],["Soviet (concept)",0.1],["German (control)",0.1],["Belarusian (relative)",-1.3],["Russia (noun)",-10.6],["Russian (core)",-10.7]]},{"id":"L1_N24822","k":"d","L":1,"N":24822,"t":"A layer-1 spelling neuron that joins W and ie across tokens","l":"This neuron detects the letters w-i-e when the tokenizer splits them, firing on an 'ie' token that follows a bare W, or on the whole token 'wie'. Its up projection recognizes ie-family tokens; its gate checks that a w came just before. Without the w, the same ie token turns…","st":"tested","i":3,"c":9,"m":[["W|IE acronym",27.8],["W|ie core",19.9],["whole wie",19.9],["Dew|ie fragment",16.0],["long w-word",13.6],["hyphen between",5.7],["control",0.0],["intact word",-0.1],["K|ie no w",-0.3],["-ie name",-1.0]]},{"id":"L20_N10159","k":"d","L":20,"N":10159,"t":"“Phew, that was close,” and its opposite, “it would prove fatal”","l":"Fires when a text says something bad almost happened: “a close one,” “dodged a bullet,” “narrowly averted.” It reads the verdict more than the event: a shark swimming past barely registers until someone says “we were lucky,” and “That was a close one” fires even after a quiet…","st":"tested","h":"../notes/L20_N10159/","i":7,"c":9,"m":[["close one",36.8],["dodged a bullet",35.2],["could've been worse",24.5],["picnic spared",22.5],["shark, no verdict",2.31],["delay proved fatal",-9.94],["would prove fatal",-14.5]]},{"id":"L20_N11305","k":"d","L":20,"N":11305,"t":"How much a body has aged or accumulated, peaking on twin paradox","l":"This neuron fires on the word that says how much one body has lived through or accumulated. Its peak is the twin-paradox outcome: how much a relativistic traveler or flown clock 'has aged', around +30. A strong halo covers lifetime tallies without relativity, such as a tram's…","st":"tested","i":8,"c":7,"m":[["twin at 0.9c",5.3],["ISS / Kelly",4.6],["flown clock",3.0],["near-light astronaut",2.4],["sailor analog",1.2],["tram tally",0.6],["ordinary aging",0.5],["nonsense",0.1],["no lifetime",-0.2]]},{"id":"L20_N11819","k":"d","L":20,"N":11819,"t":"An Esperanto detector that completes 'povas' when context is thin","l":"L20/N11819 is an Esperanto detector. It fires weakly on almost any Esperanto word piece and strongly on the word starts ' pov' (povas) and ' k' (kaj) once the context is Esperanto. Look-alike auxiliary languages such as Ido trigger it too, Italian and Spanish do not, and code…","st":"tested","i":8,"c":8,"m":[["Esperanto",63.5],["label only",47.2],["Ido",47.2],["Interlingua",45.2],["Spanish",18.6],["Italian",17.4],["nonsense",14.6],["English",13.3],["code",-24.2]]},{"id":"L20_N12396","k":"d","L":20,"N":12396,"t":"A neuron for 'its closest equivalent in another language'","l":"This neuron marks the point where a text maps something onto its counterpart in another system: another language, a programming language, or a unit system. It fires on words like 'equivalent' or 'translation', and fires hardest when the match is said to be missing or only…","st":"tested","i":7,"c":8,"m":[["gap + English",30.9],["approximate English",24.4],["gap + Italian",23.9],["plain English",19.8],["gap, no language",18.2],["plain Greek",11.9],["math, one system",10.2],["code system",8.8],["nonsense control",7.0],["unit system",5.8]]},{"id":"L20_N5517","k":"d","L":20,"N":5517,"t":"Fires negative when 'among others' or 'for example' recasts a list","l":"This neuron mostly fires negative. It does so on the first word of a phrase that recasts a finished list of nouns as a sample of a larger set: ' among' in ', among others', ' example' in ', for example', ' name' in ', to name a few'. It needs the list: the same words after a…","st":"tested","i":5,"c":8,"m":[["open-list comma",0.6],["1 item",-1.7],["and others",-3.4],["etc",-3.6],["2 items",-5.3],["clause list",-6.8],["inter alia",-8.4],["to name a few",-8.8],["for example (list)",-9.9],["among (list)",-13.7]]},{"id":"L20_N579","k":"d","L":20,"N":579,"t":"The ' while' neuron: fires on the word, scaled by the preceding predicate","l":"This neuron fires on the conjunction ' while' and almost nothing else. The word alone opens it; the clause before sets how hard. A predicate about a state, a maintained condition or a changing quantity ('the project stays on track while') gives roughly twice the response of one…","st":"tested","i":4,"c":9,"m":[["state + while",16.9],["quantity + while",9.4],["whilst",7.8],["object + while",6.8],[", and while",4.8],["even as",1.8],["plain and",0.3],["whereas",-0.1],["post-posed though",-0.1],["stacked though",-0.8]]},{"id":"L20_N9156","k":"d","L":20,"N":9156,"t":"A neuron for 'not that one': ruling out a word's other meaning","l":"This neuron fires when a writer rules out a reading of a name or term they just used, as in 'Raphael (not the Ninja Turtle)' or 'not to be confused with NM, the abbreviation for New Mexico'. It is about the word, not the world: saying 'Donatello is not the Ninja Turtle' as a…","st":"tested","i":9,"c":8,"m":[["famous namesake",28.8],["person vs unit",26.9],["planet/god",24.5],["acronym",23.8],["ambiguous word",21.4],["invented name",19.1],["no rival",17.4],["referent only",16.9],["plain object",13.2]]},{"id":"L21_N1131","k":"d","L":21,"N":1131,"t":"A neuron for names under judgment, negative on their glosses","l":"This neuron fires on the naming word itself ('named', 'name', misnomer's 'nom') when the text takes a stance on that name: beautiful, apt, cute, wrong, or a misnomer. Neutral naming barely moves it, and so do evaluative words that aren't about a name. Its negative side sits on a…","st":"tested","i":7,"c":8,"m":[["aesthetic",17.9],["aesthetic",16.1],["aptness",14.2],["aptness",12.2],["fault",12.2],["fault",9.0],["bare",4.6],["neutral",3.0],["neutral",1.2],["neutral",0.6]]},{"id":"L21_N20005","k":"d","L":21,"N":20005,"t":"A verdict detector for ideas that cannot go together","l":"This neuron fires on the word that rules on whether two ideas can go together: 'oxymoron', 'contradiction in terms', 'mutually exclusive', 'antithetical to'. It fires hardest when the verdict names an outright clash, with oxymoron near the top, and when its subject is a pair of…","st":"tested","i":6,"c":8,"m":[["concept pair",16.1],["compatible concepts",15.8],["quoted terms",12.4],["non-clash pair",10.3],["everyday things",9.0],["classic clash",8.8],["nonsense",6.9],["objects",3.1],["testimony",2.4]]},{"id":"L21_N3141","k":"d","L":21,"N":3141,"t":"A detector for already-paid costs weighed in a decision","l":"This neuron marks a cost that has already been paid at the point where a text weighs it in a decision. Two cues multiply: a decision frame and past expenditure. 'The money already spent on the plant is' fires about four times as strongly as money to be spent next year. It…","st":"tested","i":5,"c":7,"m":[["already spent",33.8],["sunk cost",31.0],["past effort",16.4],["fixed cost",12.8],["future cost",11.4],["math constant",8.2],["negligible time",5.4],["non-cost item",4.0],["no cue",2.2]]},{"id":"L21_N3827","k":"d","L":21,"N":3827,"t":"A neuron for the moment a symbol's sound is compared","l":"This neuron fires at the relation word ('like' after 'sounds', 发音, 'the digit') when a sentence is about to tie a symbol's form to something else. The symbol may be a number, word or letter, and the form may be how it sounds, how it looks or what it contains. Numerals as subject…","st":"tested","i":7,"c":6,"m":[["letter",27.4],["Chinese numeral",26.2],["digit",24.8],["numeral",23.9],["word (custom)",22.4],["word pun",18.5],["non-symbol",15.4],["nonsense",10.3],["object",6.8]]},{"id":"L21_N5036","k":"d","L":21,"N":5036,"t":"Fires when a repeated word returns as 'some more'","l":"This neuron fires at the end of a re-mention that presents something already said as a further instance: 'test some more' after 'test', or 'bigger lies' after 'lies'. Both parts are needed. The same word has to return, and a marker such as 'some more', 'more' or 'bigger' has to…","st":"tested","i":5,"c":8,"m":[["same verb",27.9],["same stem",27.1],["same + 'then'",26.6],["other earlier verb",23.8],["synonym",23.4],["new verb",20.2],["nonsense",18.9],["unrelated verb",16.2]]},{"id":"L21_N6268","k":"d","L":21,"N":6268,"t":"Finding out who your real friends are","l":"Fires when a text asks whether someone’s friends are the real thing: “hard times show you who your real friends are,” fair-weather friends, admirers who desert a man when his fortune is gone. Everyday friends, like coffee with a friend, stay near zero. It works across many…","st":"tested","h":"../notes/L21_N6268/","i":8,"c":6,"m":[["real friends are",54.2],["real allies",39.2],["friends deserted",28.6],["fair-weather friends",26.6],["true colors",24.2],["courage deserted",3.31],["coffee with friend",0.19],["fake news",-0.01]]},{"id":"L21_N6687","k":"d","L":21,"N":6687,"t":"An imagined-choice slot neuron, strongest when the choice is superpowers","l":"This neuron responds when a text invites someone to pick from an open, imagined menu, as in 'if you could have any superpower for one day, which would it be?' Superpowers push it hardest. Mundane options and bare what-ifs push it less. The same fanciful words stated as plain…","st":"tested","i":6,"c":6,"m":[["superpower pick",23.8],["magic pick",23.1],["would-you-rather",19.4],["icebreaker classic",17.0],["plain conditional",14.4],["mundane pick",12.6],["planet in frame",9.4],["no chooser",4.4],["neutral control",0.5],["flat sci-fi",-1.7]]},{"id":"L22_N10263","k":"d","L":22,"N":10263,"t":"Presented as standing for something else: bywords, examples, glosses","l":"This neuron fires when a text presents an expression, name or object as standing for something else: a byword, an example, a standard or a gloss. It peaks on the role word and the meaning that follows, not on the idiom itself, and strongest when that meaning is abstract or…","st":"tested","i":7,"c":8,"m":[["absurd idiom",19.9],["myth",18.4],["eponym",14.3],["practical use",12.1],["word",11.9],["test standard",11.5],["plain object",9.8],["nonsense",9.6],["control",8.3],["technical term",6.7]]},{"id":"L22_N13620","k":"d","L":22,"N":13620,"t":"Who gets the credit: author versus everyone else who handles a work","l":"This neuron reads a work's credit line and asks one question: did the party about to be named write the work, or only handle it? Publishers, distributors, contributors and any role label followed by a colon push it positive. The word author, in any language and any form, pushes…","st":"tested","i":5,"c":7,"m":[["publisher",7.0],["distributor",6.5],["contributor",6.1],["editor",5.0],["no-role control",4.9],["illustrator",4.4],["translator",2.4],["writer",-2.1],["co-author",-4.8],["author",-8.6]]},{"id":"L22_N14804","k":"d","L":22,"N":14804,"t":"The moment someone speaks up and nobody believes them","l":"Fires at the moment in a story when someone reports what they saw and nobody believes them: Cassandra, the boy who cried wolf, the Altamira cave. It peaks on the doubted person, the “her” in “nobody believed her.” Flip it to “everybody believed her” and it drops from about +37…","st":"tested","h":"../notes/L22_N14804/","i":8,"c":8,"m":[["Altamira cave",48.8],["truthful warning",48.5],["lying warning",41.8],["Cassandra",37.0],["not taken seriously",32.5],["everybody believed",15.0],["jury doubts",7.09],["claim unverified",-3.89]]},{"id":"L22_N19087","k":"d","L":22,"N":19087,"t":"A neuron that keeps a scientist-cast story running","l":"This neuron tracks story state. It switches on once a physicist, engineer, mathematician or other quantitative professional becomes a character in a narrative, and it stays on for the rest of the story. It is strongest for the classic scientist-joke trio but needs neither a joke…","st":"tested","i":7,"c":7,"m":[["physicist",32.0],["engineer",29.9],["programmer",29.1],["economist",26.9],["statistician",25.6],["lawyer (halo)",15.0],["blonde (joke butt)",14.5],["baker",11.2],["'a man'",7.9],["nonsense",4.8]]},{"id":"L22_N23842","k":"d","L":22,"N":23842,"t":"Sentences, numbers and rules that point back at themselves","l":"Lights up when text points back at itself: “this sentence is false,” a number whose digits count its own digits, a rule that includes itself. Swap the target and it goes quiet: “an example of itself” reads +34, “an example of inequality” +0.3. It also runs warm on liar and…","st":"tested","h":"../notes/L22_N23842/","i":9,"c":8,"m":[["self-describing number",57.2],["rule includes itself",41.8],["self-accusing knave",34.2],["example of itself",34.0],["liar sentence",24.1],["he is lying",6.56],["example of inequality",0.31],["cat washed itself",0.22]]},{"id":"L23_N22344","k":"d","L":23,"N":22344,"t":"A neuron that anticipates 'if you can call it that' hedges","l":"This neuron fires when the model expects a writer to question whether a word fits, as in 'the office, if you can call it an office'. It peaks on the naming verb, before the doubted label appears, when a conditional frame meets vocabulary about naming or applicability. It works…","st":"tested","i":7,"c":8,"m":[["if can call",41.0],["right word",36.2],["term applied",36.2],["such as it",24.0],["frame only",14.8],["right size",14.2],["name only",13.8],["if you will",6.8],["plain naming",2.0],["no cue",0.5]]},{"id":"L23_N22844","k":"d","L":23,"N":22844,"t":"A decoy detector that fires hardest when you're solving a problem","l":"This neuron fires on words that name a false lead: decoy, red herring, distractor, trap, or phrases like 'to throw you off' and 'for misdirection'. The word alone is enough, but the neuron fires about three times harder when a reader is trying to solve something, such as a math…","st":"tested","i":7,"c":8,"m":[["stage magic",26.2],["exam options",26.1],["phishing",25.2],["math problem",21.5],["detective novel",18.6],["animal eyespot",17.6],["enemy tanks",16.8],["grocery list",15.8],["nonsense",15.1],["duck hunting",7.7]]},{"id":"L23_N2419","k":"d","L":23,"N":2419,"t":"A word detector for 'composite', firing negative on first mention","l":"This neuron is a detector for the word composite. Whenever the token composite, or a translation like 复合 or zusammengesetzt, appears for the first time in a text, the neuron swings strongly negative, with formal math exposition ('composite functions') giving the deepest values.…","st":"tested","i":4,"c":7,"m":[["steel (control)",0.0],["hybrid (paraphrase)",-0.1],["combined",-2.0],["composed",-2.5],["German",-5.3],["Spanish",-5.3],["compound",-8.2],["composites",-8.4],["composite (core)",-11.1],["Chinese",-14.7]]},{"id":"L23_N4226","k":"d","L":23,"N":4226,"t":"A qualifier dial: 'near but not' fires, 'locked, or even' goes negative","l":"This neuron fires on the word that qualifies a graded statement: the 'not' in 'close to but not touching', the 'or' in 'slow or even halt'. It needs both a graded antecedent (near, similar, reduce) and a qualifying connective; either alone does nothing. When the antecedent is…","st":"tested","i":6,"c":6,"m":[["close to (core)",15.5],["almost (approx.)",13.8],["near (core)",12.2],["nearly (approx.)",12.2],["far from (halo)",10.1],["adjacent (weak halo)",7.8],["pressed against (closed)",4.9],["above (neutral)",2.8],["on (no cue)",0.7],["locked onto (closed)",-0.6]]},{"id":"L23_N8187","k":"d","L":23,"N":8187,"t":"A neuron for big endings that arrive without fanfare","l":"L23/N8187 fires on the word that describes how a big ending, discovery or surrender arrived. It fires hardest when the expected spectacle is denied: 'without fanfare', 'not with a bang', 'anticlimactic', 'no eureka moment'. It tracks the figurative sense, so literal trumpet…","st":"tested","i":8,"c":8,"m":[["without fanfare",22.5],["anticlimactic",19.5],["not with bang",18.6],["no ceremony",16.9],["literal fanfare",12.8],["with a bang",12.5],["quietly",8.8],["threshold control",1.5],["date only",0.8]]},{"id":"L24_N14165","k":"d","L":24,"N":14165,"t":"A neuron for gut feelings you can't explain","l":"This neuron fires when someone says they can't put their own reasons into words. It peaks at the end of clauses like 'can't say exactly why' or 'couldn't put a finger on it'. A preceding felt impression, such as sensed danger or unease, roughly doubles the response. The neuron…","st":"tested","i":8,"c":8,"m":[["bad feeling",45.8],["sensed danger",45.5],["felt uneasy",43.2],["warm feeling",30.4],["preference",27.4],["nonsense",19.5],["action",18.0],["bodily event",15.6],["neutral act",14.8]]},{"id":"L24_N17855","k":"d","L":24,"N":17855,"t":"Announcing how a problem will be solved, before solving it","l":"This neuron fires while a write-up is still saying how it will tackle a task, not yet tackling it. Its peaks land on the glue words of an approach sentence such as 'We can approach the problem using properties of ... and', and on the colon after 'the given problem'. Naming an…","st":"tested","i":6,"c":6,"m":[["problem",14.6],["given problem",13.2],["question",12.4],["equation",10.4],["exercise",10.1],["proof",8.5],["recipe",7.2],["coding task",5.5],["negotiation",5.3],["no-task control",0.8]]},{"id":"L24_N17907","k":"d","L":24,"N":17907,"t":"'Has to be seen': experience the reader still needs","l":"This neuron fires when a text says something must be seen, heard or experienced firsthand, and the reader has not done so yet. The clearest case is 'has to be seen to be believed', which peaks on 'seen'. The effect adds up: a perception verb in a necessity or 'unless you've'…","st":"tested","i":7,"c":8,"m":[["seen (core)",59.8],["experienced",51.2],["must be seen",48.2],["recommend",47.2],["heard",41.0],["non-perceptual",21.8],["mundane need",21.8],["no frame",13.7],["already seen",11.8]]},{"id":"L24_N18824","k":"d","L":24,"N":18824,"t":"It's not the stars moving, it's us: a frame-correction neuron","l":"This neuron fires when a text corrects an appearance by moving it onto the observer's side. The strongest case is apparent motion reassigned to the Earth, the camera or the reference frame, such as 'it is really the Earth that rotates' or 'leave the camera fixed and move the…","st":"tested","i":9,"c":8,"m":[["Earth (core)",24.6],["viewer's lens",16.9],["sky (no swap)",15.8],["camera",15.2],["observer frame",14.1],["reference value",13.7],["culprit",13.7],["blame agent",13.7],["nonsense",13.7]]},{"id":"L24_N19393","k":"d","L":24,"N":19393,"t":"Seeing a thing against its surroundings: a figure-ground comparison neuron","l":"This neuron fires when text sets a thing's look against its surroundings: a moth matching the bark, a lava flow brighter than the terrain, hidden text the color of the page. Contrast counts as well as matching, but matching wins when the text says the thing is hard to see.…","st":"tested","i":7,"c":6,"m":[["peer object",40.5],["match ground",39.5],["blend into",38.0],["contrast ground",27.4],["≈ background",25.1],["non-visual",19.1],["non-appearance",19.1],["no comparison",5.7],["control",5.7]]},{"id":"L24_N2499","k":"d","L":24,"N":2499,"t":"Succeeded where others failed: a rival's failed attempt in where/when clauses","l":"This neuron fires when a 'where' or 'when' clause sets another party's attempt against a main-clause outcome, as in 'succeeded where others had failed'. It fires most strongly when that rival failed, reaching about +55. The activity builds step by step: the link word after a…","st":"tested","i":7,"c":7,"m":[["when + others",42.2],["where + rival",38.2],["where + all others",32.5],["where, no rival",23.5],["own past failure",21.1],["locative where",19.1],["nobody else",7.8],["'but', no link",7.3],["no rival control",0.4]]},{"id":"L24_N25245","k":"d","L":24,"N":25245,"t":"Don't tell me, I want to find out myself","l":"This neuron responds to text about holding back a reveal, such as an answer, a plot, or an outcome, from the person who would otherwise learn it. Both kinds of text drive it: a listener's 'please don't tell me' and a writer's 'I won't give the plot away'. It peaks when a wish to…","st":"tested","i":7,"c":8,"m":[["novel ending",75.0],["spoilers",74.0],["puzzle answer",62.5],["nonsense",60.5],["no object",59.8],["eagle nest",51.8],["password",46.5],["time of day",38.5],["biopsy",27.0]]},{"id":"L24_N9633","k":"d","L":24,"N":9633,"t":"The 'around' in 'felt around': a neuron for searching by touch","l":"This neuron fires on the word that turns 'feel' into a search by hand: the 'around' in 'felt around', the 'way' in 'felt her way', the 'along' in 'feel along the wall'. It fires most strongly when 'feel' and a search particle occur together. Other groping verbs reach about half…","st":"tested","i":6,"c":7,"m":[["felt along",43.8],["felt around",40.5],["felt her way",33.5],["felt inside",27.6],["groped (no feel)",25.2],["felt below",24.0],["touch, no search",15.8],["emotional feel",13.3],["looking verb",0.7],["no-cue control",-0.2]]},{"id":"L25_N16371","k":"d","L":25,"N":16371,"t":"Luck that keeps repeating (+) versus outcomes decreed by fate (−)","l":"This neuron fires when text claims that luck follows a recurring pattern: Murphy's Law, 'just my luck', 'this only happens to me', or a losing streak that makes a win feel 'due'. It fires whether the luck is good or bad. A regularity with an ordinary cause leaves it silent. Its…","st":"tested","i":8,"c":6,"m":[["Murphy's law",34.8],["only me",26.6],["just my luck",24.9],["every time",20.2],["one-off luck",13.8],["coincidence",5.8],["no-cue control",3.4],["plain cause",3.2],["fate",-8.5],["God's plan",-10.5]]},{"id":"L25_N24147","k":"d","L":25,"N":24147,"t":"Confidence words that land after ignorance: a Dunning-Kruger detector","l":"This neuron fires on a word of confidence or certainty once the text has already established that the subject knows little. It is the Dunning-Kruger pairing, with or without the name. 'The least informed are the most confident' peaks around +50 at 'confident'. With 'best…","st":"tested","i":9,"c":6,"m":[["least informed",50.8],["unstudied pundits",47.8],["know nothing",46.2],["self-rating",42.5],["least skilled",28.2],["best informed",15.7],["generic people",13.9],["experts",13.2],["nonsense",8.5],["no-gap control",7.6]]},{"id":"L25_N5878","k":"d","L":25,"N":5878,"t":"Suggested options on offer versus 'the possibility exists'","l":"This neuron tracks candidate, not-yet-real things, and its sign depends on syntax. It fires positively on words that label listed items as options on offer, such as 'suggested activities' or 'suggestions for further reading' in book and course blurbs. It fires negatively on the…","st":"tested","i":5,"c":7,"m":[["suggestions for",21.2],["suggested",17.6],["possible",14.8],["proposed",13.8],["recommended",8.9],["alternative",4.5],["optional",4.4],["practical",1.4],["additional",0.8],["nonsense control",0.4]]},{"id":"L25_N6255","k":"d","L":25,"N":6255,"t":"A Monty Hall neuron that peaks when the host's knowledge is stated","l":"This neuron tracks whether the text is inside a Monty Hall-style scene: a prize hidden among closed options, a player's pick, and a second party who can open one. It builds as the scene is assembled and peaks where the informed host is about to act, especially at 'who knows'…","st":"tested","i":9,"c":8,"m":[["doors + car",59.8],["cups swap",58.5],["partial scene",57.0],["envelopes swap",56.2],["name only",53.8],["no scene",50.0],["urn cousin",45.8],["epistemic puzzle",44.8],["knower, no prize",42.0]]},{"id":"L25_N9741","k":"d","L":25,"N":9741,"t":"A 'schedule' word detector that pushes back on its calendar neighbours","l":"This neuron fires on the word 'schedule' in any spelling, sense or part of speech, strongest as a verb. Meaning is irrelevant: tax schedules and thread schedulers fire like appointments. Its negative side is the word's own neighbourhood: 'The meeting is set for Friday' scores -5…","st":"tested","i":4,"c":7,"m":[["misspelt",9.1],["scheduled (core)",5.9],["re- prefix",3.1],["synonym",3.1],["synonym",2.3],["control",-1.0],["set (fixing)",-2.9],["confirmed",-4.1],["calendar object",-4.3],["fixed (fixing)",-6.2]]},{"id":"L26_N13778","k":"d","L":26,"N":13778,"t":"A neuron that anticipates the placebo arm in medical trials","l":"This neuron fires in medical trial text at the point where the control arm or the expectation explanation is about to be named: on 'a' in 'the other group gets a' fake pill, and on 'they' in 'simply because they' expect to. The word 'placebo' alone, belief-caused improvement…","st":"tested","i":7,"c":8,"m":[["blinded trial",44.0],["drug trial",41.5],["agronomy",38.2],["placebo word",37.2],["animal sham",35.8],["belief effect",32.2],["dummy device",23.9],["nonsense",20.2],["A/B test",20.1],["no control",18.4]]},{"id":"L26_N17125","k":"d","L":26,"N":17125,"t":"A spelling neuron: bare hy fragments positive, the word debate negative","l":"This neuron reads letters, not meaning. It fires when the current token is exactly the fragment hy, or a lone y finishing a piece that ends in h, and it stays on for the next token. Whole words containing hy stay quiet. The same gate opens for the word debate and its deb-/doub-…","st":"tested","i":3,"c":7,"m":[["glued t+Hy",13.0],["bare hy",7.8],["thy halo",6.8],["Why halo",2.2],["by halo",0.6],["synonym control",0.5],["whole hy word",0.3],["no cue",0.0],["debt",-9.6],["debate",-24.2]]},{"id":"L26_N17433","k":"d","L":26,"N":17433,"t":"Little did they know: a neuron for narrated knowledge gaps","l":"This neuron fires where a story opens a gap between what someone knew then and what was about to happen. The strongest cue is an ignorance idiom about a specific narrated person, such as 'Little did I know,' or 'I had no idea that'. A retrospective 'would change my life forever'…","st":"tested","i":8,"c":7,"m":[["had no idea",40.5],["could not know",36.5],["little did I know",31.2],["looking back",27.8],["knew (no gap)",4.4],["generic you",4.3],["forgot",3.2],["plain forecast",0.7],["nonsense",0.2]]},{"id":"L26_N20030","k":"d","L":26,"N":20030,"t":"A neuron for a sports ball whose flight path bends","l":"L26/N20030 tracks a sports ball in flight and fires hardest when the text says its path bends: 'swerve', 'curve', 'bend' or 'dip', or the direction phrase right after. Three roughly additive cues build it: a ball, a competitive-sport scene, and a deviating path. A ball that…","st":"tested","i":6,"c":8,"m":[["nonsense control",31.6],["ball (core)",30.8],["football",29.4],["puck",29.4],["household object",27.5],["shuttlecock",26.9],["frisbee",26.6],["projectile",26.0],["kicked can",25.8]]},{"id":"L26_N22963","k":"d","L":26,"N":22963,"t":"A cross-lingual detector for the word turbo/turbine","l":"This neuron recognizes one word: turbo, or turbine. It fires as soon as the model knows that word is present, whether it is written in English, Russian, Chinese, Japanese or Korean, and whatever it refers to (engine, software, film). Turbulence gets a weaker version. A bare…","st":"tested","i":5,"c":8,"m":[["Russian турбина",41.0],["Chinese 涡轮",37.0],["turbine",35.0],["Turbo (software)",34.2],["turbo",31.2],["turbulence",19.9],["turban",17.8],["turmoil",9.1],["propeller (meaning)",2.1],["windmill (meaning)",0.1]]},{"id":"L26_N8504","k":"d","L":26,"N":8504,"t":"A tab-stop detector for code continuations aligned under an open paren","l":"This neuron fires on the whitespace that starts a wrapped line of code, but only when that whitespace is a whole number of tab stops wide and it lands the text roughly under the open paren of the call, def or import above. Deep kernel-style tab runs fire hardest. A separate…","st":"tested","i":5,"c":6,"m":[["2 tabs, exact",12.8],["16 sp, exact",11.0],["17 sp, one long",7.4],["8 sp, one under",6.7],["15 sp, one short",4.8],["24 sp, one over",3.1],["32 sp, two over",1.9],["1 sp, control",0.4],["12 sp, off-grid",-0.4],["4 sp, plain indent",-0.5]]},{"id":"L26_N8815","k":"d","L":26,"N":8815,"t":"Change verbs inside a time-travel frame: tampering with the past","l":"This neuron fires when someone acts on events that have already happened: a time traveler changing the past, stopping an assassination, killing a grandfather, or being told the timeline cannot be altered. A time-travel frame sets a modest baseline. A change verb inside that…","st":"tested","i":8,"c":7,"m":[["time machine",39.2],["contingent history",35.8],["fate",13.9],["trip to past",7.2],["regret",6.8],["nonsense",2.7],["mundane",2.3],["rational agent",1.8],["factual",1.3]]},{"id":"L27_N13773","k":"d","L":27,"N":13773,"t":"Onsets of backwards-spelled words, with negative word-order reversal","l":"This neuron marks where each character-reversed word begins, whether the model is reading backwards text or writing it on request. It strengthens as a list establishes the reversal pattern, peaking around +15. The last letters of a reversed word go negative, and so do reordered…","st":"tested","i":8,"c":6,"m":[["reversed word",11.9],["final capital",11.9],["reversed name",9.6],["scrambled letters",8.8],["nonsense",7.1],["word-order reversal",6.5],["forward word",0.5],["spaced reversal",-6.0],["fused equation",-6.4]]},{"id":"L27_N15531","k":"d","L":27,"N":15531,"t":"A neuron for programs pausing until you press any key","l":"This neuron tracks a program that stops and waits for an empty go-ahead: 'press any key', getch, Console.ReadKey, cv2.waitKey. The pause-until-the-user frame alone brings it partway, about 30. The content-free signal word ' key' roughly doubles that, peaking near 100. Fuller…","st":"tested","i":6,"c":8,"m":[["any key",94.5],["a key",77.0],["Enter to continue",63.2],["space bar",62.0],["click OK",48.8],["read output",43.2],["typed number",38.2],["command",35.8],["password",29.0]]},{"id":"L27_N17377","k":"d","L":27,"N":17377,"t":"An indefinite act of analysis offered as what a text contains","l":"This neuron fires when a text is said to contain, or a judgment to rest on, an act of analysis phrased indefinitely: 'the report provides a comparison of', 'the ruling depends on a careful examination of'. The same phrase as something a person does fires about a third as…","st":"tested","i":5,"c":7,"m":[["consideration",14.1],["comparison (core)",9.9],["examination",6.9],["definite article",5.8],["list (neutral)",3.3],["summary (weaker)",2.6],["concrete noun",1.9],["partitive",1.9],["quantifier",-2.1],["item adjective",-6.2]]},{"id":"L27_N20837","k":"d","L":27,"N":20837,"t":"Only the tip: observed samples framed as part of more","l":"This neuron fires when text calls something seen, reported or found only a small visible part of a larger hidden whole. 'Tip of the iceberg' is the stock example. Two things drive it and partly overlap. One is the observed-sample frame ('revealed so far... only the'). The other…","st":"tested","i":6,"c":8,"m":[["uncovered frauds",40.5],["known cases",35.8],["revealed so far",35.0],["few examples",27.5],["bare subject",18.5],["species found",13.9],["neutral control",6.3],["non-sample subject",4.2],["nonsense",3.8],["plain quantity",1.9]]},{"id":"L27_N25489","k":"d","L":27,"N":25489,"t":"The song you can't get out of your head","l":"This neuron fires for earworms: a jingle stuck in your head, a kid who won't stop singing Baby Shark, humming a tune you barely heard. It likes songs that live inside people. Concerts, albums and music on the radio get almost nothing.","st":"tested","h":"../notes/L27_N25489/","i":8,"c":7,"m":[["song stuck in head",67.5],["kid humming it",58.5],["Chinese earworm",51.2],["Baby Shark again",40.8],["children's song",16.5],["song on radio",5.75],["Beethoven concert",0.01],["named recording",-1.91]]},{"id":"L27_N8726","k":"d","L":27,"N":8726,"t":"A neuron for the fourth wall between a work and its audience","l":"This neuron tracks the line between a performed or written work and the people receiving it. It fires hardest when that line is named: the 'fourth wall' idiom in any language, whether the wall is broken, kept or forbidden. It also fires when a narrator, actor or author crosses…","st":"tested","i":7,"c":8,"m":[["idiom (ES)",73.0],["idiom (EN)",54.5],["frame break",41.2],["literal wall",41.2],["metafiction",33.5],["real performer",21.2],["neutral control",20.9],["marketing",18.0],["museum (neg)",11.9]]},{"id":"L28_N10548","k":"d","L":28,"N":10548,"t":"A plastic-shopping-bag waste neuron that peaks before the decomposition figure","l":"This neuron tracks the plastic shopping bag as disposable waste. Naming the bag alone gives a moderate response. A throwaway-and-persistence frame gives a partial response on its own, even for bottles or cups. When the two combine in bag-ban facts or 'takes up to 1,000 years to…","st":"tested","i":6,"c":7,"m":[["plastic bag",65.5],["carrier bag",64.0],["straw",54.2],["paper bag",45.0],["tote bag",39.5],["no object",38.5],["water bottle",37.2],["coffee cup",27.8],["sleeping bag",22.4],["granite",5.0]]},{"id":"L28_N1380","k":"d","L":28,"N":1380,"t":"Is the website down? A negative neuron for online service status","l":"This neuron works mainly through its negative side. It goes strongly negative on words that say whether an online service is working: Twitter is down, the site crashed, the apps are back online. It responds to the topic of service status, not to failure itself, so up and not…","st":"tested","i":6,"c":8,"m":[["power",-2.9],["sun (control)",-2.9],["elevator",-5.4],["ATM",-9.6],["phone lines",-9.6],["nonsense",-9.6],["servers",-19.8],["website",-27.8],["app",-38.0],["Twitter",-45.2]]},{"id":"L28_N15222","k":"d","L":28,"N":15222,"t":"Garages where things get built, not where cars get parked","l":"This neuron fires on the room where something gets made or started: Apple's garage, a founder's basement, a dorm room. The word \"garage\" alone does little: a parked car scores near 8, two founders push it past 50, and the split holds in eight languages. Turned up, it sends a…","st":"tested","h":"../notes/L28_N15222/","i":8,"c":8,"m":[["Apple's garaje",61.8],["Jobs's garage",60.0],["Jobs's 车库",57.0],["first speakers",55.5],["parents' basement",48.2],["beer brewing",41.0],["parked car",7.72],["mechanic's garage",2.44]]},{"id":"L28_N16278","k":"d","L":28,"N":16278,"t":"A cross-lingual detector for office the place, not office the post","l":"This neuron responds to the word for an office as a place, in any language, and to almost nothing else. Translations like Büro and ufficio fire as hard as English office, but the same languages' words for a held position stay silent. In English, where one word carries both…","st":"tested","i":6,"c":8,"m":[["novel modifier",25.8],["office",24.6],["Italian ufficio",17.9],["Russian офис",17.5],["German Büro",11.6],["compound",9.3],["bureau",5.9],["set phrase",2.5],["room control",0.2],["workplace",0.1]]},{"id":"L28_N18006","k":"d","L":28,"N":18006,"t":"Time that flies, and time that drags","l":"Fires when a text talks about how fast time feels: the hour that flew by, the minute that felt like an hour, tips for making the workday go faster. It works across languages. Clock time, timetables and flying birds leave it near zero. Real changes in time's speed, as in…","st":"tested","h":"../notes/L28_N18006/","i":7,"c":8,"m":[["having fun",113],["Russian",109],["workday faster",76.0],["felt like an hour",59.2],["hour flew by",41.0],["relativity",18.5],["clock on wall",4.38],["bird flew by",0.089]]},{"id":"L28_N18801","k":"d","L":28,"N":18801,"t":"Where the sea pulls back: a tsunami-motion neuron","l":"This neuron tracks the sea moving sideways as one mass. It fires hardest on the verb when water pulls back before a tsunami, rushes in, or travels as a wave with a speed and height, peaking near +58. Tides, floods and seiches fire about half as strongly. A crowd or army…","st":"tested","i":7,"c":8,"m":[["tide",46.2],["bathtub",38.0],["ocean",37.5],["harbor water",36.8],["sea",36.0],["floodwater",32.0],["lake",17.1],["crowd",10.8],["nonsense",6.7],["army",5.9]]},{"id":"L28_N19665","k":"d","L":28,"N":19665,"t":"A neuron that names beliefs that make themselves come true","l":"This neuron fires when text is about to name a belief as the cause of its own outcome. It peaks on 'self-fulfilling prophecy', or a translation, after a story in which an expectation made itself come true. It often peaks on ' self' before the phrase is finished. Explicit loop…","st":"tested","i":7,"c":7,"m":[["explicit loop",66.0],["refuted forecast",39.2],["foretold fate",38.8],["label loop",38.5],["self-defeating",29.5],["bank run",18.5],["nonsense",16.8],["weather",15.4],["no prediction",8.4]]},{"id":"L28_N24107","k":"d","L":28,"N":24107,"t":"'It doesn't take a ___': the denied-competence slot neuron","l":"This neuron fires on the word that opens a credential slot inside a denied competence requirement: 'It doesn't take a ___', 'You don't need to be a ___', 'I'm no ___', 'even a ___ could'. The denial is essential. After the same setup, 'It takes a' barely registers. Argued…","st":"tested","i":8,"c":7,"m":[["doesn't take a",44.0],["don't need a",43.0],["I'm no",26.1],["anyone with",15.1],["even a",13.6],["plain statement",9.5],["takes a (no denial)",1.3],["need a (no denial)",0.6],["no-frame control",0.1],["hedged 'perhaps a'",-0.5]]},{"id":"L28_N3639","k":"d","L":28,"N":3639,"t":"A news outlet's open choice, and the 'content' it produces","l":"This neuron has two signs, and both concern the news media. It turns mildly positive at the word right after a newspaper or news app takes up a choice or offer that is still open, as in 'a newspaper offers [three]'. It turns strongly negative, about ten times deeper, where media…","st":"tested","i":4,"c":6,"m":[["magazine",2.7],["newspaper",1.9],["news site",0.9],["bank control",-0.3],["publisher",-0.6],["blog",-0.8],["nonsense",-1.0],["bakery control",-1.2],["news app",-1.3],["radio",-1.6]]},{"id":"L28_N4359","k":"d","L":28,"N":4359,"t":"A neuron that recognises the word optical in any script, sense-blind","l":"This neuron fires on the token that completes the word for optics, whether English 'optical', Spanish 'ópticas', Russian 'оптический', Chinese 光学 or letters spelled out one by one. It reads the word, not the meaning: political 'optics' fires nearly as hard, while lens, light and…","st":"tested","i":5,"c":8,"m":[["optical (core)",57.8],["Russian оптический",44.5],["bare 'opt' fragment",40.8],["Spanish óptica",40.5],["Chinese 光学",40.2],["plain adjective + optics",30.0],["lexicalised compound",20.1],["light word, no 'opt'",13.6],["rival: optimal",13.4],["no-cue control",-0.1]]},{"id":"L28_N7008","k":"d","L":28,"N":7008,"t":"Fires when an expected day is moved; negative on plain arrival or waiting","l":"This neuron fires when text says an expected future event will happen at a different time than expected. Its strongest trigger is a verb moving a named time point: 'bring forward the day when' or 'postpone the moment'. Direction matters little there. Without a time noun it…","st":"tested","i":6,"c":7,"m":[["hasten",25.4],["bring forward",23.6],["postpone",16.5],["speed up",15.8],["delay",15.4],["no shift",4.8],["wait for",3.7],["celebrate",3.4],["predict",3.1],["nonsense control",1.8]]},{"id":"L28_N854","k":"d","L":28,"N":854,"t":"Dust in the air reads negative, cloud droplets forming read positive","l":"This neuron tracks fine matter in a medium and signs it. Dust, pollen, flour, soot or silt present in a scene push it negative, twice as hard when the particles hang in air or water, and the peak lands on the word after the particle noun. Water vapor condensing into cloud…","st":"tested","i":6,"c":6,"m":[["gas accretion",3.5],["cohering silk",1.5],["control",-0.2],["condensing vapor",-0.7],["cloud droplets",-1.8],["bubbles",-2.5],["discrete objects",-5.2],["soot",-7.2],["pollen",-14.4],["dust (core)",-23.0]]},{"id":"L28_N8779","k":"d","L":28,"N":8779,"t":"Open-ended capability, about to be pinned to one limit","l":"This neuron fires when text has said a capability is open-ended and is about to name the one thing that bounds it: 'limited only by …', 'the only limit is …', 'the sky's the limit'. The peak comes before the factor is named. It fires as hard for 'limited only by compute time' as…","st":"tested","i":6,"c":7,"m":[["limited only",81.0],["bare bound",81.0],["restrictive only",81.0],["solely",68.0],["only limited",66.0],["determined",57.5],["shaped",20.1],["hindrance",15.1],["control",9.3],["nothing excluded",8.1]]},{"id":"L28_N9500","k":"d","L":28,"N":9500,"t":"A neuron for learned human senses, peaking on human echolocation","l":"This neuron fires when text names a sense someone learned through practice. Human echolocation is the prototype: 'echolocation' in a sentence about a person learning to click and hear echoes drives the peak, whether or not the person is blind or succeeds. Weaker halos cover…","st":"tested","i":8,"c":6,"m":[["sighted human",29.0],["blind human",25.2],["deaf human",23.2],["child",15.5],["nonsense",11.8],["robot",9.8],["dog",8.5],["bat (innate)",6.2],["dolphin",4.9]]},{"id":"L29_N14034","k":"d","L":29,"N":14034,"t":"Fires just before 'the first' in milestone and record reports","l":"This neuron marks the moment just before a record claim. When text reads like a milestone report (a dated chronicle, 'made history', a known world-first, or a study 'believed to be the first'), it fires on the verb or article right before 'the first', peaking near +80. The…","st":"tested","i":7,"c":8,"m":[["world first",85.5],["dated milestone",76.0],["mundane record",30.5],["role change",12.2],["research prose",9.8],["famous name, off-topic",9.4],["no-cue control",1.0],["likely 'first', no record",0.8],["nonsense",0.6]]},{"id":"L29_N18164","k":"d","L":29,"N":18164,"t":"A li detector: the 礼 character, and 'ritual' after Confucius","l":"This neuron detects li (礼), the Confucian idea of ritual and propriety. The standalone 礼 character fires in almost any context, even inside a subway station name. English words like ritual, rites and propriety fire only after a Confucian cue: 'corrected by ritual' gives +32…","st":"tested","i":7,"c":8,"m":[["Xunzi",38.0],["Legalist",37.5],["Confucius",37.2],["modern wedding",37.2],["Zhou rites",35.5],["Mencius",34.2],["nonsense",32.8],["Japanese court",29.2],["Durkheim",5.1]]},{"id":"L29_N21111","k":"d","L":29,"N":21111,"t":"'The other side' as a far place where something awaits","l":"This neuron fires on ' other' in 'on the other (side/end/line)' when the phrase names a separate far place where something waits or is heard: the life after a tunnel of grief, the voice on a phone line, the stranger behind a screen. A passage in the preceding text pushes it near…","st":"tested","i":7,"c":8,"m":[["hardship passage",81.5],["phone line",76.0],["awaiting frame",64.5],["unknown beyond",32.2],["plain far place",21.9],["bare locative",10.2],["discourse 'other hand'",2.2],["object face",-1.1],["equation side",-1.3]]},{"id":"L29_N21165","k":"d","L":29,"N":21165,"t":"Household things left idle, peaking on 'collecting' before 'dust'","l":"This neuron fires when a portable household thing (an exercise bike, puzzle books, camera lenses) has been described as sitting idle in a room or on a shelf. It peaks on the verb just before 'dust', as in 'ended up in the spare room collecting dust'. The resting frame alone…","st":"tested","i":6,"c":7,"m":[["garage",76.0],["spare room",70.0],["attic",64.5],["closet",56.8],["museum",47.0],["storage unit",46.5],["abstract",45.5],["no-place",42.5],["trash",28.1]]},{"id":"L29_N3332","k":"d","L":29,"N":3332,"t":"Their, there, they're: a neuron for pronoun homophones named as words","l":"This neuron fires when English pronoun homophones are treated as words rather than used as words. Examples are their/there/they're and, more weakly, its/it's and your/you're. Activation builds across a list and peaks on its last member. A bare list is enough to trigger it, and…","st":"tested","i":5,"c":7,"m":[["their",39.0],["they're",36.2],["it's",28.6],["two",23.5],["there",21.8],["affect",21.4],["his (pronoun)",20.8],["your",15.6],["whether",10.2],["nonsense",6.2]]},{"id":"L29_N9871","k":"d","L":29,"N":9871,"t":"As the crow flies versus walking the route leg by leg","l":"This neuron tracks how a text gets from one place to another. It is most active when a direct separation between two points has just been stated and a straight-line framing is expected, peaking on 'the' in 'as the' before 'crow' appears. It goes negative when a route is traced…","st":"tested","i":7,"c":6,"m":[["invented place",40.5],["real town pair",37.5],["by road",35.2],["leg-by-leg walk",34.8],["wayfinding",32.8],["polygon perimeter",31.4],["no-place control",30.4],["grid",27.5],["astronomy",23.0]]},{"id":"L2_N13982","k":"d","L":2,"N":13982,"t":"A letter-level n-vowel-n reader: nin negative, nan/nen/non positive","l":"This neuron watches letters, not words. Whenever a token closes a run of n, vowel, n, it fires, and the vowel sets the sign. An i in the middle (Knin, Lenin, a spelled-out K n i n) drives it strongly negative, about three times harder than the other vowels push it positive…","st":"tested","i":4,"c":8,"m":[["a-chain 3rd n",9.2],["e-chain 3rd n",7.6],["o-chain 3rd n",4.6],["consonant + in",0.5],["no n control",0.0],["one n only",-0.2],["a before in",-16.2],["whole-token nin",-18.9],["spelled out",-19.4],["n-i-n core",-26.5]]},{"id":"L2_N15786","k":"d","L":2,"N":15786,"t":"A letter-by-letter F-then-O detector that reports matches as negatives","l":"This neuron watches text being spelled letter by letter and fires when the current token completes F followed by O, as in U.F.O., L|FO or Ru|fo. It reports the match on its negative side, dropping to about -30. A half match, a spelled O with no F before it or a spelled F not…","st":"tested","i":3,"c":10,"m":[["spelled O, no F",1.0],["spelled F, no O",1.0],["lexical word",0.1],["V control",-1.6],["bare pair",-5.8],["ph digraph",-11.2],["Cyrillic",-19.2],["other F-O acronym",-21.5],["whole token",-24.5],["dotted U.F.O.",-32.2]]},{"id":"L2_N21338","k":"d","L":2,"N":21338,"t":"An assembl- spelling tag with an inert 'Question' artifact","l":"This early-layer neuron detects a spelling, not a meaning. It fires on pieces of the assemble stem (assemble, unassembled, Spanish asamblea), but stays silent on 'assembly' and on synonyms like 'gather'. Lookalikes such as ensemble get a faint echo. Its corpus maximum, about +52…","st":"tested","i":3,"c":8,"m":[["assemble",14.8],["re-assemble",10.2],["unassembled",9.2],["Spanish stem",7.8],["resemble",0.2],["ensemble",0.1],["synonym",0.0],["-assembly noun",0.0],["Question (late)",-0.0],["nonsense",-0.0]]},{"id":"L2_N5966","k":"d","L":2,"N":5966,"t":"A Greek letter-chain detector for οργαν-, plus a stray 'Question' artifact","l":"Qwen spells Greek words one letter at a time, and this layer-2 neuron follows a single chain: ο-ρ-γ-α-ν. Activation builds with each letter and peaks (about +30) on the closing ν of οργανισμός, όργανο or οργάνωση, in any case or sense. The neuron then nudges the model toward ι…","st":"tested","i":6,"c":8,"m":[["all caps",32.8],["core οργαν",32.5],["ε- near-miss",19.0],["α- misspelling",17.2],["Cyrillic root",17.1],["ορ_αν, no γ",12.6],["English whole-token",6.1],["γ, no ορ",3.7],["no-chain control",-0.0],["other -αν",-0.4]]},{"id":"L2_N8195","k":"d","L":2,"N":8195,"t":"A string detector for the name Srinivasa, in any script","l":"This neuron reads one letter string, not a topic. It fires when subword pieces spell the Indian name Srinivasa or Srivastava, on the piece that completes the word, in Latin, Cyrillic or Devanagari. Two identifier bigrams, S2CID and UniProt, ride the same detector. Context does…","st":"tested","i":4,"c":5,"m":[["Srinivasa (core)",47.0],["all caps",37.0],["Cyrillic",27.1],["Sreenivasa (halo)",27.1],["Srivastava (core)",26.0],["Sriram (sri- onset)",10.8],["space-broken",5.4],["other Indian name",1.9],["Ramanujan (control)",0.0],["English name (control)",0.0]]},{"id":"L2_N9582","k":"d","L":2,"N":9582,"t":"A spelling gate for St-et + son/oscope: Stetson and stethoscope","l":"L2/N9582 is a spelling detector for Stetson and stethoscope. It fires on the completion piece ('son' or 'oscope') when that piece follows an St+et(h) stem. Meaning plays no part: hats, the university, people, the instrument and made-up blends like Stethson all hit the same…","st":"tested","i":5,"c":10,"m":[["Stetson (core)",34.0],["crossed nonword",34.0],["stethoscope (core)",27.0],["stem variant",23.6],["suffix variant",20.4],["Sk stem",10.3],["other -oscope",4.2],["no-cue control",0.0],["-son surname",-0.7],["wrong stem",-0.8]]},{"id":"L30_N11765","k":"d","L":30,"N":11765,"t":"The 'as' after 'grow': naming the facet of change","l":"This neuron goes strongly negative on a postposed 'as' (or Italian 'come', Spanish 'como') that names the facet in which a person changed: 'helped me grow as', 'sono cresciuta come'. Its deepest values, around -75, come from reflective growth text. Judgment verbs ('fails as')…","st":"tested","i":6,"c":6,"m":[["act as",4.1],["hired",0.4],["role",0.0],["nonsense",-0.5],["excel (judge)",-3.7],["regard",-4.4],["fail (judge)",-4.8],["change",-24.5],["mature",-36.0],["grow (core)",-37.8]]},{"id":"L30_N6476","k":"d","L":30,"N":6476,"t":"A signed slot: named later stage versus remedied open future","l":"This neuron fires on the word that opens a later time, before that time is named, in sentences where something done now matters later. Its sign depends on what kind of later the sentence sets up. A specific later stage ('succeed after [graduation]') gives a modest positive. A…","st":"tested","i":6,"c":8,"m":[["succeed after",9.2],["later",5.4],["avoid suits after",4.0],["neutral in",0.3],["next year",-5.2],["years after",-12.7],["plain forecast in",-12.7],["avoid mistakes in",-24.8],["learn feedback in",-39.5],["moving forward",-39.8]]},{"id":"L30_N8699","k":"d","L":30,"N":8699,"t":"The pending verdict when a rival makes something look lesser","l":"This neuron fires at the moment one thing's appearance is about to be judged against a rival it has been set beside: 'her handwriting makes mine look', 'pales in', 'puts X to', and the same shadow and dimming idioms in German, Italian, French, and Dutch. It fires before the…","st":"tested","i":8,"c":8,"m":[["makes mine",45.2],["makes yours",44.0],["puts to (idiom)",40.8],["compared with",13.4],["next to",6.6],["makes me",6.4],["plain comparison",6.2],["third party",2.7],["resembles",0.2],["no rival",0.1]]},{"id":"L31_N10626","k":"d","L":31,"N":10626,"t":"A neuron for a body lying still inside an MRI scanner","l":"This neuron tracks one scene: a body, whether volunteer, patient or dog, lying still inside an MRI or fMRI scanner. It fires hardest on 'lie still', 'inside' and the article just before the scanner is named, reaching about +140. Naming MRI without a body inside, or talking about…","st":"tested","i":7,"c":8,"m":[["fMRI scanner",104],["PET scanner",81.5],["MRI machine",78.0],["CT scanner",70.5],["ultrasound",68.0],["hyperbaric tube",66.5],["airport scanner",66.5],["kennel",66.5],["X-ray",66.5],["nonsense",66.5]]},{"id":"L31_N13367","k":"d","L":31,"N":13367,"t":"A neuron for the ace up your sleeve, across languages","l":"This neuron fires when a text commits to a stock idiom for a concealed extra resource: 'a trick up one's sleeve', 'an ace in the hole', 'a rabbit out of the hat', or German 'Ass im Ärmel'. It peaks on the token that locks the idiom in, not on the noun. Looser paraphrases such as…","st":"tested","i":6,"c":8,"m":[["ace up",235],["tricks up",161],["ace in hole",135],["rabbit/hat",122],["trump card",82.0],["plan B",56.0],["secret weapon",45.2],["literal noun",45.2],["control",45.2],["frame only",45.2]]},{"id":"L31_N1552","k":"d","L":31,"N":1552,"t":"Drawing a survey sample: stratified allocation up, study analysis down","l":"This neuron fires when a text specifies how to draw a sample from a real population: which sampling method, and how many units from each subgroup. Stratified sampling is its strongest case, in English, French or Chinese, and it promotes 'strat' as the next word when the design…","st":"tested","i":5,"c":8,"m":[["simple random",57.0],["margin of error",53.5],["stratified rock",46.8],["分层抽样",46.5],["proportional allocation",46.2],["no cue",45.0],["cluster",43.2],["stratified",41.5],["control group",40.2],["MANOVA variables",32.8]]},{"id":"L31_N16641","k":"d","L":31,"N":16641,"t":"Costumes as a figure of speech","l":"Fires when something is described as wearing a costume, as a figure of speech: the wolf in sheep's clothing in five languages, racism in disguise, old wine in new bottles. Neither hidden danger nor a real swap matters much: a sheep in wolf's clothing gets +100.5, and even a…","st":"tested","h":"../notes/L31_N16641/","i":8,"c":8,"m":[["wolf in sheep's clothing",130],["sheep in wolf's clothing",100],["old wine, new bottles",63.5],["tiger in a tiger's fur",50.5],["tax hike in disguise",34.8],["Trojan horse",9.81],["tax hike, plain",0.81],["man in a hoodie",-23.2]]},{"id":"L31_N23558","k":"d","L":31,"N":23558,"t":"Solve-for-a-value versus defined-as-a-sum at the word 'of'","l":"This neuron tracks which kind of math sentence is being built and reports it at ' of'. It fires positive when an equation has just been posed and the sentence asks for a value ('If 4^(x-1)·4^(3x)=4^15, what is the value of'), and more so the more math the condition holds. It…","st":"tested","i":5,"c":4,"m":[["exponent eq",33.0],["exponent, plain",17.5],["linear eq",6.9],["assignment",1.2],["painting value",0.2],["non-math cond",-0.3],["aggregate, problem",-0.5],["divisibility rule",-24.5],["perfect-number def",-42.0]]},{"id":"L31_N3689","k":"d","L":31,"N":3689,"t":"Is the feeling returned? Plus a 'one step ahead' negative","l":"This neuron fires on the word saying whether one person's attitude toward another is returned, shared or reciprocated: 'return', 'share', 'corresponder', 'разделены', 'unrequited'. It responds most when the thing returned is a feeling. That feeling can be love or hatred, and the…","st":"tested","i":6,"c":7,"m":[["love",85.0],["admiration",68.0],["favour",66.5],["hatred",65.0],["respect",62.2],["greetings",48.2],["calls",24.6],["nonsense",20.4],["letters",10.3],["money",1.2]]},{"id":"L31_N3975","k":"d","L":31,"N":3975,"t":"The slot before the program you need to open a named file","l":"Once a text names a file by its format and turns to what the reader needs to open it, this neuron fires on the word just before the program is named: 'you will need the', 'can be opened with'. It peaks near +60 for PDFs, archives and video; floppy disks and slides work more…","st":"tested","i":6,"c":7,"m":[[".mov video",48.5],["PDF",44.5],["ZIP archive",40.5],["floppy disk",30.4],["hardcover",26.1],["35mm slides",25.2],["no format (control)",23.1],["executable (run)",22.9],["Docker image",9.1],["cookies (control)",9.0]]},{"id":"L31_N542","k":"d","L":31,"N":542,"t":"Cleaner animals picking food off hosts, anticipating the meal","l":"This neuron fires when a small animal takes food or parasites off another creature or surface. Oxpeckers picking ticks from a rhino and cleaner wrasse at a grouper's gills are the textbook cases. Each cue adds to the signal: a stereotyped cleaner name, a host, a picking verb and…","st":"tested","i":8,"c":6,"m":[["oxpecker+host",87.0],["cleaner fish",78.0],["plover+croc",53.2],["name, no host",52.2],["host, generic bird",47.8],["carcass feeder",34.5],["generic forager",24.8],["human",15.6],["bridge crossing",4.6],["tethered dog",-9.2]]},{"id":"L31_N8472","k":"d","L":31,"N":8472,"t":"A neuron that measures how Titanic-like a sea voyage is","l":"This neuron measures how closely a passage matches the Titanic story: a large steam-era passenger ship, at sea, at night, heading into ice. The ingredients multiply rather than add, and no ship name is needed; a fictional liner in an ice field scores highest. The signal peaks on…","st":"tested","i":9,"c":6,"m":[["steamer in ice",67.5],["liner in ice",50.2],["Titanic",46.5],["liner, calm",46.5],["liner, reefs",38.5],["no vehicle",36.2],["train",35.0],["fishing boat, ice",30.4],["airliner, ice",29.4],["Grand Canal",28.2]]},{"id":"L32_N10049","k":"d","L":32,"N":10049,"t":"A pride, a murder, a flock: waiting for the word for a group of animals","l":"This neuron fires right before the model names a group of animals, as in \"a group of owls is called a...\", and on the group words themselves, most of all a lions' \"pride\". The name needn't be whimsical: a flock of sheep counts almost as much as a business of ferrets. A group of…","st":"tested","h":"../notes/L32_N10049/","i":8,"c":7,"m":[["lions' prides",56.2],["group of ferrets",49.0],["group of sheep",44.2],["baby kangaroo",23.1],["group of ships",11.2],["group of students",0.96],["pocket chart",-5.91]]},{"id":"L32_N13110","k":"d","L":32,"N":13110,"t":"One gate, two signs: wild-stock inventories up, sustainability meta-talk down","l":"This neuron fires when a state counts or shields a standing wild resource: a national forest inventory, a wildlife survey, a fishing ban, a protection project, strongest when the state is China or Korea and weak for Western agencies. Planting trees, cleaning pollution or…","st":"tested","i":5,"c":6,"m":[["fishery survey",14.9],["wildlife survey",14.5],["China forest (core)",13.7],["forest, no nation",11.8],["Canada forest",11.8],["soil (halo)",7.5],["water quality (halo)",2.2],["census (control)",-0.7],["traffic (control)",-0.7],["transdisciplinary (neg)",-18.6]]},{"id":"L32_N15857","k":"d","L":32,"N":15857,"t":"One held body in a mechanics problem versus velocity-update rules","l":"This neuron is a signed axis. It fires positively when text sets up an idealized textbook mechanics problem: one body sliding off a sphere, a mass dropped on a spring, or the abstract version, setting an energy functional stationary. It peaks on the word just before the quantity…","st":"tested","i":6,"c":6,"m":[["gas molecules",29.8],["puck (core)",29.6],["loaf mimic",29.6],["particle swarm",29.1],["spring block",28.8],["bead",28.1],["tourist mimic",25.9],["flock",24.1],["nonsense",23.0]]},{"id":"L32_N16883","k":"d","L":32,"N":16883,"t":"A leech neuron that fires where the bloodsucking is named","l":"This neuron tracks a clause-local state: a leech, or a described bloodsucking worm, is the thing being used. The leech word switches the state on, but the peak comes a few words later, on the slot just before the use is named ('used by doctors to'). That slot is the causal part.…","st":"tested","i":8,"c":8,"m":[["leeches",149],["medicinal leech",144],["leech, unnamed",121],["lamprey",82.5],["cupping",55.0],["tick",44.5],["snail",42.2],["tapeworm",40.0],["towel control",10.8],["nonsense",7.0]]},{"id":"L32_N19087","k":"d","L":32,"N":19087,"t":"Velcro and heat-pipe detector that fires where the fact comes next","l":"This neuron picks out two unrelated kinds of device. The first is the Velcro fastener, reached by any name, by the phrase 'hook-and-loop', or by its inventor. The second is heat-moving hardware that uses a working fluid, such as cryocoolers, heat pipes and vapor chambers. It…","st":"tested","i":5,"c":5,"m":[["hook-and-loop",91.5],["Velcro",87.5],["cryocooler",70.0],["gecko halo",52.0],["heat pipe",43.8],["German name",25.1],["zipper",15.8],["plain loop",4.0],["fridge control",0.6],["nonsense",-0.5]]},{"id":"L32_N19670","k":"d","L":32,"N":19670,"t":"Halloween neuron fires just before the trick-or-treat haul is named","l":"This neuron tracks Halloween trick-or-treating. Once the scene is set, it stays moderately active, and it peaks on the word just before the text names what the children collect or what they should do about it: 'lots of' before candy, or a 'Don't' in a safety list. Each piece of…","st":"tested","i":7,"c":6,"m":[["guising",128],["Halloween",110],["trick-or-treat",106],["door-to-door Tuesday",65.5],["St. Martin's",54.2],["Oct 31 date",48.5],["Easter",43.2],["nonsense",33.0],["Tuesday",31.2],["Christmas",18.5]]},{"id":"L32_N25129","k":"d","L":32,"N":25129,"t":"A neuron for everyone chipping into the same hat","l":"This neuron fires when several people each put something into one shared container whose contents will be drawn, counted, or given away: names into a hat, dollars into a hat for a funeral, the church plate, a poker kitty. Activation builds through the scene and peaks on the…","st":"tested","i":7,"c":8,"m":[["collection plate",50.2],["hat",49.5],["ballot box",39.2],["jar",36.0],["coffee can",34.5],["basket",30.6],["mailbox",16.5],["recycling bin",11.9],["no container",8.6],["cathedral",2.4]]},{"id":"L32_N5531","k":"d","L":32,"N":5531,"t":"Bug under a glass: invertebrates in containers versus vertebrate anatomy","l":"This neuron tracks a small live invertebrate, such as a spider, beetle or moth, in a concrete scene. It fires hardest when a person confines the creature, peaking at the container word in phrases like 'trapped it under a glass' (a spider goes from about 29 to 42). Snails and…","st":"tested","i":7,"c":7,"m":[["moth",46.8],["wasp",46.8],["beetle",42.2],["spider (core)",42.0],["snail (halo)",31.0],["crab (halo)",26.8],["mouse (vertebrate)",18.8],["bird",18.5],["frog",17.5],["no animal",9.4]]},{"id":"L32_N8251","k":"d","L":32,"N":8251,"t":"A groundhog-name detector that makes 'Day' less likely next","l":"This neuron recognizes one word family: groundhog, woodchuck, marmot and Punxsutawney. It fires on the token that completes the name, whether the text means the animal, the film, the idiom or a company. February or Candlemas context earlier in the passage raises it toward its…","st":"tested","i":6,"c":8,"m":[["groundhog",145],["woodchuck",118],["Punxsutawney",108],["marmot",90.5],["ground squirrel",75.0],["hedgehog",37.5],["prairie dog",34.8],["rabbit",20.2],["no animal",18.6],["badger (omen)",15.5]]},{"id":"L32_N8734","k":"d","L":32,"N":8734,"t":"A Baroque-style detector that turns negative on painted colours","l":"This neuron tracks the Baroque as a named art, architecture or music style. It fires on the style word in any European language, peaks near 100 when text defines the term, and stays high on the words that follow. Other period styles reach a third to a half of that. Inside the…","st":"tested","i":5,"c":8,"m":[["Baroque",68.5],["barocco (It.)",57.2],["Roman Baroque",47.8],["Rococo",32.5],["Gothic",28.4],["Renaissance",27.4],["Brutalism",25.8],["artists, no word",14.9],["control",4.4],["grey hue",-17.1]]},{"id":"L32_N901","k":"d","L":32,"N":901,"t":"The slot where a train's unknown length is about to be written","l":"This neuron fires, negatively, at the moment a train-crossing problem is about to write down a length that the text has not yet fixed. The space after 'A train', the plus sign in 'Total distance covered = 360 +', the 'the' in 'the length of the train' all sit near its floor of…","st":"tested","i":8,"c":8,"m":[["rope (control)",-6.2],["marching column",-8.6],["ship",-12.4],["truck",-20.4],["tunnel",-25.4],["platform",-29.9],["bridge",-30.1],["nonsense",-36.0],["train (core)",-36.8],["goods train",-37.2]]},{"id":"L33_N11266","k":"d","L":33,"N":11266,"t":"Polar light and sea-ice extremes, opposed by polar ocean heat","l":"This neuron responds when text describes something at high latitude swinging to a seasonal extreme: months of polar night or midnight sun, or sea ice reaching its annual minimum or maximum. The response shrinks as the same sentence moves toward the equator, from +74 at an…","st":"tested","i":9,"c":6,"m":[["Antarctic station",71.5],["South Pole",53.2],["Tromsø",48.0],["Oslo",28.8],["nonsense",25.8],["London",23.4],["Singapore",16.5],["Arizona",13.7],["Caribbean",13.3]]},{"id":"L33_N13890","k":"d","L":33,"N":13890,"t":"Shivaji detector: strong negative on Maratha names, weak Indian-context positive","l":"This neuron has one gate that opens on Indian-name material, and its up projection chooses the sign. The large pole is negative and lexical. It fires on the final token of Shivaji (about -40) and, more weakly, on Sambhaji, Shahaji, Bajirao and Balaji, whoever the name refers to.…","st":"tested","i":4,"c":7,"m":[["Indian surname",6.2],["Western name",0.5],["Japanese -ji",-0.1],["no name",-0.1],["same 'aji'",-2.1],["Balaji",-17.8],["Peshwa",-18.5],["honorific -ji",-19.0],["Shivaji's son",-27.2],["Shivaji",-33.5]]},{"id":"L33_N17747","k":"d","L":33,"N":17747,"t":"An umami neuron: MSG, glutamate enhancers, and savory stock","l":"This neuron tracks umami, the savory fifth taste. It fires on MSG, glutamate, ribonucleotide enhancers, yeast extract, kombu and dashi, and on stock or broth used as a savory base. It usually peaks on the slot just before the enhancer or broth word arrives. Salt, sweeteners and…","st":"tested","i":7,"c":6,"m":[["yeast extract",83.0],["ribonucleotides",80.0],["kombu",71.0],["stock",53.8],["MSG",40.2],["salt",5.2],["honey (sweet)",4.8],["lime (tangy)",4.8],["sugar",4.8],["no-food control",4.8]]},{"id":"L33_N18392","k":"d","L":33,"N":18392,"t":"Planetary geology seen from orbit, with two silent software squatters","l":"The neuron tracks passages about a rocky world's surface studied from a distance: orbital gravity maps, crater imagery, magma-ocean crust, probes crashed into the Moon. The Moon drives it hardest, with Mars and Mercury at about half. Two things switch it off: people or rovers…","st":"tested","i":7,"c":5,"m":[["far side",46.8],["Moon (core)",41.0],["Mars",27.4],["Mercury",25.1],["Afghanistan (halo)",17.6],["Vesta",12.6],["Earth",10.5],["Pluto (halo)",8.7],["no-place control",8.2],["Earth place name",4.6]]},{"id":"L33_N18511","k":"d","L":33,"N":18511,"t":"Health bars and gauges: a negative state for displayed levels","l":"This neuron goes negative when text describes a display whose fill or position stands for a quantity. Game HUD resource bars are strongest: the noun 'bar' in 'health bar' or 'hunger bars' reaches about -35. Progress bars, sliders and gauge needles come next, and the state…","st":"tested","i":6,"c":5,"m":[["control",-2.8],["no display",-4.0],["HP variable",-5.3],["other bar",-8.5],["health (no bar)",-9.9],["gauge needle",-11.5],["XP bar",-21.0],["battery bar",-23.4],["progress bar",-23.9],["health bar",-24.5]]},{"id":"L33_N21908","k":"d","L":33,"N":21908,"t":"The next regrowable body part in a list","l":"This neuron tracks lists of body parts that grow back. It fires hardest on the separator that opens the next item, as in 'can regrow its tail, legs,' (about 86). The frame matters more than the animal: mice that cannot regrow reach the same level. Salamanders and axolotls boost…","st":"tested","i":8,"c":8,"m":[["axolotl regrows",86.5],["mice cannot",83.5],["lizard",74.5],["newt regenerates",70.5],["machine regrows",68.0],["axolotl, no regrowth",59.5],["named dog",43.2],["no-cue control",2.3],["injury only",0.6],["woodwork control",-2.2]]},{"id":"L33_N2301","k":"d","L":33,"N":2301,"t":"Newton-type versus projection/proximal: a signed optimizer-family neuron","l":"In optimization text, this neuron signals which family of method is being described. Newton-type, damped, barrier and stochastic-approximation procedures push it positive: Levenberg-Marquardt, interior-point, Gauss-Newton and Robbins-Monro. The convex-analytic first-order school…","st":"tested","i":7,"c":5,"m":[["Levenberg-Marquardt",17.9],["interior-point",16.2],["Robbins-Monro",11.0],["BFGS",10.9],["Adam",6.4],["no method",4.3],["plain GD",-5.1],["mirror descent",-19.9],["Nesterov",-23.0],["proximal",-27.2]]},{"id":"L33_N4149","k":"d","L":33,"N":4149,"t":"A hanging-topic detector: bare nouns asked as questions drive it negative","l":"This neuron's real job is on its negative side. It fires hardest at the question mark that closes a bare content word used as a whole utterance, such as 'Ben?' or 'Beethoven?', and the effect spills onto the first word of the reply. The harder the noun is cut off from any verb…","st":"tested","i":7,"c":5,"m":[["wh-word + ?",0.2],["noun + verb",-0.1],["noun + period",-3.4],["alternative + ?",-4.3],["noun + comma",-6.5],["noun + !",-7.4],["ellipsis + pronoun",-8.6],["echoed name + ?",-12.0],["noun + ?",-13.9],["demonstrative + ?",-14.6]]},{"id":"L33_N583","k":"d","L":33,"N":583,"t":"Snow that reflects sunlight, with aging skin as the opposite pole","l":"This neuron responds to snow as a bright white surface that throws sunlight back. It fires most when UV or sunlight is said to reflect off snow, as in sun-safety tips, snow-glare warnings and snow albedo. It fires more weakly for albedo in general and for other reflectors such…","st":"tested","i":6,"c":7,"m":[["snowy slopes",50.8],["fresh snow",49.2],["glacier ice",45.2],["white sand",39.0],["lawn control",31.9],["water",27.9],["asphalt",27.9],["salt flats",25.6],["aged skin",21.9],["nonsense",19.8]]},{"id":"L33_N7903","k":"d","L":33,"N":7903,"t":"A neuron for 'two tired': numbers pressed into the degree slot","l":"This neuron fires on the adjective that completes a degree phrase when the word in the degree slot is not a real degree adverb. The slot word is usually a small number or 'this/that'. The bicycle pun 'It's two tired!' is its signature case. With real 'too' or 'very' the neuron…","st":"tested","i":9,"c":8,"m":[["numeral",28.0],["numeral (pun)",25.2],["demonstrative",17.1],["numeral homophone",11.8],["numeral + very",10.4],["homophone 'four'",8.9],["article control",0.8],["real 'very'",0.1],["arbitrary noun",-0.7],["real 'too'",-1.5]]},{"id":"L34_N12965","k":"d","L":34,"N":12965,"t":"Go-hand-in-hand idiom detector, negative for linked, mixed and traded","l":"This neuron reads the moment a sentence commits to two named things traveling together, as in 'go hand in hand' and its Spanish, Italian, German and Chinese versions. It responds to the fixed idiom, not to the idea of complementarity: 'are complementary' barely registers. The…","st":"tested","i":5,"c":8,"m":[["go hand",28.9],["go together",13.1],["side by side",11.4],["inseparable",7.2],["complementary",6.5],["control",-1.0],["linked",-4.9],["weighed against",-8.1],["traded for",-10.1],["mixed",-10.2]]},{"id":"L34_N13078","k":"d","L":34,"N":13078,"t":"A Tooth Fairy neuron: shed tooth, pillow swap, coin","l":"This neuron tracks the lost-tooth custom: a child's tooth comes out, goes under the pillow, and the Tooth Fairy or a local version of her leaves money. A child's shed tooth, the pillow routine and the fairy's name each raise it, and together they push it past 100. Adult…","st":"tested","i":7,"c":6,"m":[["young child",98.0],["child",90.0],["toddler",85.0],["animal",70.0],["adult",67.0],["elderly adult",57.8],["adult, trauma",53.2],["nonsense control",50.0],["clinical",48.0]]},{"id":"L34_N14483","k":"d","L":34,"N":14483,"t":"The creaky-floorboard neuron: squeaks from floors, stairs, beds and hinges","l":"This neuron tracks the creak or squeak of things people step, sit or lie on, or swing open: floors, stairs, bed frames, hinges, pipes. It fires just as hard when the sound is denied ('no squeaking'), used as a metaphor, or about to come. The same noise from knees, brakes or…","st":"tested","i":8,"c":6,"m":[["stair",36.8],["bed frame",33.5],["hinge",32.5],["brakes",24.6],["saddle",21.6],["knees",20.0],["dishwasher",12.1],["cat",12.1],["nonsense",12.1]]},{"id":"L34_N15918","k":"d","L":34,"N":15918,"t":"Invented sound spellings up, paired-word slots like 'nip and tuck' down","l":"This neuron has two poles. It fires positively inside freshly invented spellings of sounds, such as a bird's 'kwirrp' or a reactor's 'vwum', most strongly when the text frames them as sounds. A real sound word that is a single token, like 'chip', gets roughly half as much or…","st":"tested","i":6,"c":7,"m":[["coined call",19.9],["coined call 2",16.0],["real word (1 token)",8.1],["verb + object",3.7],["non-sound word",3.5],["sound verb",1.3],["punctual verb",-0.1],["bare verb",-13.1],["doublet opener",-13.8],["nip-and-tuck",-24.6]]},{"id":"L34_N21393","k":"d","L":34,"N":21393,"t":"A signed axis over 'two' words: bis( up, bipartite and biennial down","l":"This neuron fires on words built from a 'two' prefix, and the word itself fixes the sign. Chemical bis( connectors, binary code and doubling arithmetic push it positive, most strongly inside dense technical text. Bipartite, bisexual, bilingual and recurrence periods such as…","st":"tested","i":4,"c":6,"m":[["binary code",5.0],["chem bis(",3.5],["diploid halo",3.3],["bidirectional",2.7],["double",1.7],["no-two control",0.0],["chem di- control",-2.3],["bilingual",-3.7],["bipartite",-9.4],["biennial",-12.3]]},{"id":"L34_N22154","k":"d","L":34,"N":22154,"t":"A yawning-word detector, strongest when yawns are contagious","l":"This neuron is a word detector for yawning. It fires on the first token that makes a word mean 'yawn', in English, German or Spanish, and is strongest when the text talks about yawning being contagious (about +30). Boredom, drowsiness and other contagious reflexes give only weak…","st":"tested","i":4,"c":4,"m":[["contagious yawn",27.4],["German yawn",22.5],["plain yawn",20.5],["Spanish yawn",15.8],["boredom",10.8],["described, no word",7.6],["sneezing",4.8],["laughter",3.9],["aviation yaw",2.2],["control",-0.2]]},{"id":"L34_N24759","k":"d","L":34,"N":24759,"t":"A Trojan Horse topic neuron that stays on for fact recall","l":"This neuron tracks the Trojan Horse as the thing under discussion. It peaks on the word naming the horse (about +118) and stays high on the next slots where a fact about it is recalled, such as the date or the builder. The generic malware 'trojan' counts as the same entity.…","st":"tested","i":8,"c":6,"m":[["Ilium",83.0],["Troy (core)",82.5],["Trojans' city",82.5],["Athens",67.0],["Troy NY",60.2],["Carthage",55.0],["Jerusalem",47.8],["nonsense",39.5],["warehouse",34.2]]},{"id":"L34_N25535","k":"d","L":34,"N":25535,"t":"Accent-slot oddities mid-word: broken -ção, pinyin ó up; South-Asian surnames down","l":"This neuron fires on a sub-word token in the accent slot of a word the model recognizes, when that token rarely appears there. Strongest: a Portuguese -ção suffix with corrupted or stripped accents, and a tone-marked ó standing alone inside a pinyin syllable. Rare names split…","st":"tested","i":4,"c":6,"m":[["pinyin ó",29.4],["odd-piece name",0.8],["rare name halo",0.4],["mojibake -ção",-0.2],["correct -ção",-0.2],["plain word",-0.2],["common surname",-0.3],["stripped -ção",-0.4],["known-name piece",-1.1],["South Asian surname",-4.7]]},{"id":"L34_N2741","k":"d","L":34,"N":2741,"t":"One Neuron, Several Contextual Completion Families","l":"This neuron responds to several specific contextual completions: Review after Artificial Intelligence, Dart html, beryllium numbers, and selected sex and gamete phrases. Dialog fields, Mirza Ghulam Ahmad, and some natural-history completions suppress it. Matched controls show…","st":"tested","i":2,"c":2,"m":[["AI title",8.1],["Dart completion",7.2],["Sex and cell completion",6.4],["Beryllium outside chemistry prose",5.2],["Matched cell control",0.3],["Ordinary-name control",-0.1],["Matched Review control",-0.3],["Matched element control",-0.3],["Matched class-name control",-0.3],["Harmless dialog field",-5.8]]},{"id":"L34_N4693","k":"d","L":34,"N":4693,"t":"Sulfite, not sulfate: a 3-versus-4 sign across memorized frames","l":"This neuron says 'sulfite, not sulfate'. It fires hardest on the token that pins down a sulfite species: the 'ite' of sulfite or bisulfite in any language, or the 3 closing SO3 in a formula. Sulfate and SO4 are a shallow dip. The same 3-versus-4 sign reappears in a handful of…","st":"tested","i":9,"c":5,"m":[["UUID3",50.8],["SO3 formula",37.2],["sulfite word",30.5],["probe 3",23.1],["sister -ite",17.9],["sulfur(IV) halo",17.6],["other salt",-0.3],["sulfate",-2.8],["UUID4",-23.4],["R-code 4",-38.0]]},{"id":"L34_N7546","k":"d","L":34,"N":7546,"t":"Senkaku/Diaoyu dispute detector, with a faint Haitian-1791 positive side","l":"This neuron is a detector for the China–Japan dispute over the Senkaku/Diaoyu islands. It fires strongly negative (down to about −64) on the island names and on the 'the' just before an expected name. Strength fades for other East Asian island disputes and vanishes for Western…","st":"tested","i":8,"c":7,"m":[["Falklands",-2.2],["Gibraltar",-4.8],["no place",-5.2],["Japan–Russia",-10.9],["China–Vietnam",-13.7],["China–Philippines",-15.9],["Japan–Korea",-20.9],["Taiwan–Japan",-27.1],["China–Japan",-37.8]]},{"id":"L35_N10320","k":"d","L":35,"N":10320,"t":"A technical-sense reader: empirical formula, Routh, radius of convergence","l":"This neuron fires downward on a handful of technical terms, and only in their technical sense. Its strongest response is the chemistry task of deriving a formula from percent composition, with or without the words 'empirical formula'; the eponym Routh in stability analysis and…","st":"tested","i":4,"c":3,"m":[["hydrology",0.2],["control theory",-0.3],["economics",-0.4],["physics",-3.6],["nonsense",-4.8],["materials science",-5.7],["pharmacy",-11.9],["stoichiometry",-41.2],["chemistry",-41.5],["hydrocarbon analysis",-44.0]]},{"id":"L35_N12983","k":"d","L":35,"N":12983,"t":"A neuron for fire as the stolen prize of myth","l":"This neuron responds to fire as a prize in origin-of-fire legends: fire stolen, lost or won, as when Prometheus or Coyote steals it. A steal verb is enough on its own. Other verbs work only inside a myth frame. It already responds at 'stole the', before fire is named. Other…","st":"tested","i":8,"c":7,"m":[["fire",21.5],["wallet",21.2],["sacred fire",20.8],["sun",20.8],["water",20.8],["moon",20.8],["table",20.8],["cattle",12.6],["nonsense",5.1]]},{"id":"L35_N14330","k":"d","L":35,"N":14330,"t":"Brood-parasite words plus an egg-laying frame, peaking at 'lays its eggs in'","l":"This neuron fires when a text about birds reaches the egg-laying slot of a brood-parasite story: 'the cowbird lays its eggs in'. Two cues add: an egg-laying or nest frame, and a word that names brood parasitism (cuckoo in any language, cowbird, 'brood parasite'). Either alone…","st":"tested","i":8,"c":8,"m":[["coucou (French)",114],["cuckoo",112],["cowbird",103],["cuckooshrike (non-parasite)",100],["cow- control",99.5],["invented bird",83.0],["non-bird",74.0],["robin",64.5],["honeyguide (parasite)",62.5]]},{"id":"L35_N15966","k":"d","L":35,"N":15966,"t":"Milgram's neuron: firing where ordinary people are about to obey","l":"This neuron tracks Stanley Milgram and his central claim: ordinary people obey authority into harm. It fires on the surname itself and at the point in an obedience story where 'they obeyed' or 'following orders' comes next, even when no name appears. Ablating it removes the…","st":"tested","i":8,"c":6,"m":[["Milgram",92.0],["unnamed shocks",84.5],["harmless task",55.8],["Nuremberg",48.8],["Zimbardo",40.0],["Harlow",23.0],["Asch",20.9],["no cue",2.0],["other Milgram",1.6],["operant",-12.6]]},{"id":"L35_N1900","k":"d","L":35,"N":1900,"t":"A neuron for broken-heart syndrome, peaking where the name comes due","l":"This neuron tracks takotsubo, or broken-heart syndrome: the real condition in which emotional shock temporarily stuns the heart muscle. It fires strongest when the syndrome is named, given a synonym, or clearly about to be named, in many languages. The figurative 'broken heart'…","st":"tested","i":7,"c":6,"m":[["name itself",43.8],["grief",41.0],["divorce",39.2],["synonym",28.0],["physical stress",22.1],["joyful shock",16.6],["heart attack",15.0],["plain cardiology",13.6],["no-cue control",12.8]]},{"id":"L35_N22096","k":"d","L":35,"N":22096,"t":"Owning your work versus pooling it: a signed axis in edtech prose","l":"This neuron is a signed axis, not an education detector. It fires positive on the word own, owns or control once context has fixed the sense of responsibility for one's own work, in classrooms or sales teams alike. It fires negative, two to five times harder, on nouns naming…","st":"tested","i":5,"c":5,"m":[["edtech practice",6.5],["ownership (core)",5.1],["property sense",4.2],["control",2.2],["shared drive",0.9],["no-cue control",0.8],["community of practice",-0.9],["library",-1.3],["Moodle",-1.9],["ComPADRE",-10.9]]},{"id":"L35_N25473","k":"d","L":35,"N":25473,"t":"Shape of living body surfaces, opposed to properties of physical media","l":"This neuron responds to words for the shape, thickness or texture of a living body's surface: flat, thin or thick skin patches, conical bumps, a dog's coat, a cat's tubular tail. Colour and itch words barely move it, and the same words on a wall or leaf are weaker. Its stronger…","st":"tested","i":5,"c":6,"m":[["lay skin",4.8],["clinical skin",4.8],["dog",3.5],["statue",3.4],["leaf",2.4],["no surface",2.1],["plaster wall",1.8],["exhaust pipe",0.9],["physics medium",-13.2]]},{"id":"L35_N8620","k":"d","L":35,"N":8620,"t":"A neuron that anticipates coined names for dips, led by 'uncanny valley'","l":"This neuron fires when text is about to name a coined term for a dip, trough or island in a trend. The uncanny valley is by far its favorite. Island of stability, double descent and the hype-cycle trough follow. A naming frame alone, or a dip alone, gives only weak firing. It…","st":"tested","i":7,"c":8,"m":[["uncanny dip",47.5],["island of stability",44.2],["double descent",42.0],["hype trough",31.4],["post-lunch dip",26.5],["Moore's law",20.5],["happiness U",15.4],["bystander effect",8.7],["nonsense",3.2]]},{"id":"L35_N9580","k":"d","L":35,"N":9580,"t":"A weekday-date detector that fires just before the day number","l":"This neuron marks the moment a weekday-headed date is about to give its number. It fires on the bare space token right after the month in 'Monday, November 17', and it fires equally whether a day or a year follows. Without the weekday, the same slot is silent. Weaker date heads…","st":"tested","i":4,"c":8,"m":[["weekday (core)",34.8],["other weekday",32.5],["German weekday",27.5],["abbreviation",24.9],["Today",10.2],["holiday",3.5],["season",1.6],["Posted",0.5],["fake word",-0.3],["no head",-0.5]]},{"id":"L36_N10322","k":"d","L":36,"N":10322,"t":"Waiting for the Moon to cover the Sun","l":"This neuron keeps a running sense that the Sun's disk is about to be covered, with the Moon as the expected occluder. It rises over a passage and peaks on the small word just before 'total solar eclipse' would come, up to about +120. Other bodies crossing the Sun get about half…","st":"tested","i":8,"c":7,"m":[["Sun, uncovered",79.0],["Moon over Sun",67.5],["Moon over lamp",55.8],["cloud",55.0],["sunset",51.5],["lunar eclipse",41.0],["Venus transit",40.5],["balloon",40.2],["Phobos (Mars)",34.0],["no-sky control",3.9]]},{"id":"L36_N1145","k":"d","L":36,"N":1145,"t":"A neuron for illusions of sight, named or described","l":"This neuron fires on the idea of an illusion of sight. It responds most strongly to the phrases 'optical illusion' or 'visual illusion' in any language or setting, even as a proper name. A textbook comparison of the form 'these are equal, but one looks different' gets close…","st":"tested","i":6,"c":6,"m":[["visual illusion",39.8],["optical illusion",32.8],["optical trick",30.2],["cognitive idiom",27.9],["bare illusion",26.9],["design effect",20.6],["non-visual",20.0],["trompe-l'oeil",13.1],["control",-2.5],["pattern (neg)",-5.3]]},{"id":"L36_N12088","k":"d","L":36,"N":12088,"t":"A Hitchhiker's Guide neuron that lights up before the Answer","l":"This neuron tracks The Hitchhiker's Guide to the Galaxy and its joke that 42 is the Answer. It fires hardest, around +150, on the title token 'Hitchh|iker' and on the word just before the Guide names its Answer. Weaker, graded echoes appear on meaning-of-life questions…","st":"tested","i":8,"c":9,"m":[["Guide lore",140],["Guide title",131],["Adams novel",126],["Adams other",23.2],["literal hitchhiker",22.4],["meaning of life",20.4],["Lost numbers",18.9],["nonsense",16.0],["other novel",13.1],["math quiz",2.0]]},{"id":"L36_N13666","k":"d","L":36,"N":13666,"t":"Paying for self-cultivation, and against code that registers commands","l":"This neuron tracks money going to a person's growth or pastime: schooling, courses, books, lessons and hobbies. It peaks on the slot that names or leads into that object, such as 'training budget' or 'can afford to send'. A money frame or a learning topic alone barely moves it.…","st":"tested","i":5,"c":8,"m":[["courses",24.0],["books",17.0],["hobbies",14.9],["vacation (halo)",9.9],["schooling",7.5],["lessons",7.1],["jewelry (halo)",5.1],["nonsense",1.1],["necessity",1.0],["necessity",0.1]]},{"id":"L36_N13959","k":"d","L":36,"N":13959,"t":"Tabs or spaces: a neuron for choosing indentation width","l":"This neuron fires when text treats indentation as a width to choose: tabs versus spaces, or how many spaces per level, in prose, editor configs or other languages. Its strongest activation often lands on the slot just before the unit, such as 'configure your editor to convert'…","st":"tested","i":6,"c":6,"m":[["editor setting",54.2],["vim config",27.9],["indent frame",23.8],["Word tab stop",22.5],["cursor key",21.4],["unit, no frame",3.8],["no cue",0.5],["case shift",-4.0],["prosodic stress",-15.9]]},{"id":"L36_N14127","k":"d","L":36,"N":14127,"t":"A neuron for the input-type side of function signatures","l":"This neuron tracks the input side of a written signature: the list of argument types or spaces and the arrow that closes it. It fires most on 'B: V × V →' in math and on bare type lists such as 'Item Number ->' in code, and goes quiet once the codomain starts. The factors have…","st":"tested","i":6,"c":6,"m":[["scalar mult.",32.8],["nouns",28.9],["V×V (core)",27.0],["dual pairing",27.0],["three factors",27.0],["arithmetic",21.8],["single factor",17.9],["lowercase vars",17.5],["stub (negative)",5.2],["no domain",3.0]]},{"id":"L36_N14369","k":"d","L":36,"N":14369,"t":"Is it still the same one? An identity-through-change neuron","l":"This neuron tracks the identity-through-change question: after a story in which a thing's parts are replaced, copied or transformed, it fires where the text asks whether it is still the same one. The Ship of Theseus is the prototype, but toasters, teams, teleported persons and…","st":"tested","i":9,"c":8,"m":[["all parts replaced",45.5],["rebuilt from copies",37.2],["melted & recast",34.0],["repainted",14.6],["reclassified (kind)",8.8],["one part replaced",7.1],["nonsense",6.4],["no change",6.0],["just old",3.2]]},{"id":"L36_N14935","k":"d","L":36,"N":14935,"t":"A neuron that expects the rainbow's seven colours","l":"This neuron tracks the visible-spectrum colour set: the rainbow, Newton's prism, ROYGBIV and VIBGYOR. It fires hardest just before that set is named ('seven colors of the') and on indigo inside the list. Other colour systems and other fixed lists, such as weekdays and solfège…","st":"tested","i":8,"c":6,"m":[["VIBGYOR",55.2],["rainbow",29.6],["Newton prism",23.2],["primaries",18.2],["pride flag",13.2],["weekdays",12.0],["EM spectrum",9.5],["no set",-2.8],["-fold",-7.7],["Sanskrit count",-17.6]]},{"id":"L36_N15664","k":"d","L":36,"N":15664,"t":"Spoken for: goods and people pledged to a counterparty","l":"This neuron fires negative when a good or person has been assigned to a counterparty and the claim is in force but the handover is not yet done: forward contracts to buy or sell, a daughter promised in marriage, a backlog of plane orders, a recruit committed to a college. The…","st":"tested","i":9,"c":8,"m":[["delivery, no claim",3.4],["free transfer",-2.4],["plans",-9.2],["market fact",-11.0],["sold, plain",-11.5],["bare future",-12.9],["agreed",-18.8],["under contract",-19.1],["promised",-21.0],["sold, marked",-30.2]]},{"id":"L36_N17878","k":"d","L":36,"N":17878,"t":"Telescope-era Mars and Germany entering Belgium, negative on 1944","l":"This middle-layer neuron has two separate themes. The first is Mars as seen from Earth before spacecraft: Schiaparelli, Lowell, canals, oases and polar caps, peaking around +75. Rover-era Mars fires only weakly. The second is German armies entering Belgium and the West in 1914…","st":"tested","i":6,"c":6,"m":[["Lowell oases",50.2],["Schiaparelli canals",50.0],["Wells Martians",36.5],["Belgium 1914",32.2],["Belgium 1940",30.5],["Russia 1812",5.2],["rover Mars",1.0],["no-cue control",0.4],["Hungary 1944",-7.7],["Arnhem 1944",-8.9]]},{"id":"L36_N18301","k":"d","L":36,"N":18301,"t":"Detects the second letter of a freshly spoken spelling","l":"This neuron marks the moment a text begins spelling a word aloud, letter by letter. It fires on the second letter, where a lone capital becomes a spelling run, and then fades. It responds most to the first spelling in a context, when the letters could start a real word and…","st":"tested","i":7,"c":7,"m":[["core spelling",72.5],["repeat spelling",72.5],["plausible onset",64.0],["odd onset",33.8],["single letter",22.0],["phonetic slashes",20.0],["period-separated",19.9],["impossible onset",18.4],["digits",17.8],["no spelling",4.5]]},{"id":"L36_N1854","k":"d","L":36,"N":1854,"t":"A slot for greeting and celebration rituals, peaking before 'five'","l":"This neuron fires where text expects a customary contact ritual: a high five, fist bump, handshake, cheek kiss or hug done as a greeting or celebration. It anticipates the gesture rather than detecting its word. 'In Brazil, the standard greeting between friends is the' already…","st":"tested","i":7,"c":7,"m":[["cheek-kiss custom",33.2],["Brazil custom",30.1],["violent contact",28.2],["sports win",25.9],["bow custom",23.2],["nonsense",19.8],["verbal phone",18.6],["kids",18.0],["verbal email",7.6]]},{"id":"L36_N20720","k":"d","L":36,"N":20720,"t":"Better in every group, worse overall: a Simpson's paradox neuron","l":"Fires for Simpson's paradox: two sides compared on a rate, then split into groups (each department, every airport, both halves), where the total can flip. It is already high at the split, before the twist, works across languages, ignores storks-and-babies and Monty Hall, and…","st":"tested","h":"../notes/L36_N20720/","i":9,"c":7,"m":[["Berkeley departments",81.5],["same, in Chinese",76.0],["every airport, flipped",75.5],["named outright",51.2],["WWII bullet holes",45.5],["rate, no split",27.5],["storks and babies",2.77],["Monty Hall",1.12]]},{"id":"L36_N24132","k":"d","L":36,"N":24132,"t":"Storybook pirates and buried treasure, peaking where X marks the spot","l":"This neuron tracks the storybook pirate world: Golden-Age pirates, their emblems such as the Jolly Roger and skull and crossbones, and the buried-treasure map where X marks the spot. The treasure schema alone is enough, since a chest buried on an island fires it with no pirate…","st":"tested","i":7,"c":6,"m":[["modern piracy",57.8],["outlaw+emblem",57.0],["named pirate",50.8],["emblems only",49.0],["pirate emblem",48.0],["no-pirate burial",40.8],["no-scene control",37.5],["modern criminal",37.2],["geocache",35.5],["hoard control",33.8]]},{"id":"L36_N24142","k":"d","L":36,"N":24142,"t":"A neuron for the shock wave of airbursts and explosions","l":"This neuron tracks the air-blast of a sudden, huge release of energy: the shock wave from a meteor airburst, eruption or detonation, and the windows and trees it knocks down far away. The signal builds over the sentence and peaks on 'shock wave', in any language. Naming Tunguska…","st":"tested","i":7,"c":7,"m":[["Tunguska",64.5],["Siberia",61.0],["Chelyabinsk",57.8],["Beirut",53.2],["Texas control",52.5],["generic suburb",52.2],["Hiroshima",48.0],["fireworks",47.5],["nonsense",47.5]]},{"id":"L36_N24324","k":"d","L":36,"N":24324,"t":"Where the placenta sits: a previa neuron in Qwen3-32B","l":"This neuron tracks statements about where the placenta sits in the uterus. It fires hardest on the word 'previa' and on the determiner right before the landmark is named, whether the placenta covers the os or lies clear of it. Other pregnancy structures at the cervix inherit…","st":"tested","i":5,"c":6,"m":[["previa itself",57.2],["placenta (core)",41.8],["cord (halo)",21.5],["membranes (halo)",18.6],["gallstone",18.0],["fibroid",13.8],["pterygium (eye)",12.9],["truck (non-medical)",7.2],["Toyota control",6.2],["nonsense control",6.2]]},{"id":"L36_N24563","k":"d","L":36,"N":24563,"t":"Predicts a what-it-meant-to-them predicate after a fronted human topic","l":"This neuron listens for sentences that say what an event meant to a group of people. Once a human topic has been fronted by any device, 'For the villagers,' 'As for the fishermen,' 'From the boatmen's perspective,' it swings strongly negative on the clause verb before any…","st":"tested","i":6,"c":7,"m":[["temporal 'for'",-0.5],["reportative",-0.6],["no topic",-0.8],["region topic",-2.7],["for + group",-3.8],["to + group",-4.0],["non-human topic",-4.9],["named individual",-5.0],["perspective frame",-6.8],["as for + group",-7.0]]},{"id":"L36_N5796","k":"d","L":36,"N":5796,"t":"An open set U as the domain of a map","l":"This neuron recognizes the standard setup of analysis and topology: an open region declared and then used as the domain of a map, as in 'Let U ⊂ ℝⁿ be open, f: U → ℝ'. The letter U alone carries much of the signal. Intervals fire nearly as strongly, while closed, compact or…","st":"tested","i":5,"c":7,"m":[["open Banach",51.5],["open ball",36.8],["open set ℝⁿ",36.5],["discrete grid",36.5],["bare letter",35.0],["open interval",32.2],["closed ball",32.0],["mock place",25.4],["compact",21.6],["no region",9.2]]},{"id":"L36_N6351","k":"d","L":36,"N":6351,"t":"What kind of two-place object is being put on this set?","l":"When formal text equips a set with a two-place object, this neuron reports what kind of object it is. Relations drive it sharply negative: the symbol ~, the words relation or equivalence, the product S × S, and the phrase Zeroth Law, which reaches about -60. Valued objects such…","st":"tested","i":6,"c":6,"m":[["distance",10.9],["norm",6.2],["group structure",5.4],["metric",5.3],["topology (control)",3.0],["partial order",-5.0],["nonsense",-14.1],["binary operation",-15.3],["equivalence",-27.5],["relation",-35.5]]},{"id":"L36_N8234","k":"d","L":36,"N":8234,"t":"A neuron for trolley-problem setups: a moving vehicle and a lever","l":"This neuron recognizes a trolley-problem setup: a vehicle heading toward something while someone, usually 'you', can still divert it. Naming 'the trolley problem' is enough, and so is the bare scenario with no victims at all. The word 'trolley' alone does nothing. Lifeboats…","st":"tested","i":8,"c":5,"m":[["trolley 5 vs 1",58.0],["named",56.8],["no victims",55.0],["driverless car",41.0],["shopping trolley",30.0],["Schrödinger",27.9],["lifeboat",22.1],["Monty Hall",19.8],["no cue",10.2],["everyday dilemma",5.0]]},{"id":"L37_N10659","k":"d","L":37,"N":10659,"t":"Record schema slot axis: surrogate id negative, data columns positive","l":"Inside a table or entity definition, this neuron tracks which slot the text is in. The surrogate-id slot and anything that fills it (auto-increment, SERIAL, @GeneratedValue, even 'an id column that is' before any keyword) drive it strongly negative, down to about -47. Ordinary…","st":"tested","i":7,"c":8,"m":[["data column",4.6],["text data column",3.7],["numeric data column",2.5],["nonsense control",2.0],["foreign key",-5.5],["UUID, no generation",-10.8],["natural key",-12.2],["id slot",-18.2],["surrogate key",-26.2],["auto-increment lexeme",-39.5]]},{"id":"L37_N11163","k":"d","L":37,"N":11163,"t":"The Sisyphus neuron: predicting the boulder and the climb","l":"This neuron fires for the Sisyphus picture: a person straining a heavy stone up a slope. It lights up on the name itself, on the stone and 'up' in the scene, and even on the word just before the boulder when context makes it the obvious next word. Other mythic punishments and…","st":"tested","i":9,"c":8,"m":[["unnamed king",58.0],["Prometheus",53.2],["condemned",49.8],["Sisyphus",40.8],["Atlas",36.5],["nonsense",27.1],["bulldozer",26.9],["gym",26.4],["driver",9.6],["stroller",4.3]]},{"id":"L37_N11511","k":"d","L":37,"N":11511,"t":"The rock too heavy for God to lift","l":"Fires on the omnipotence puzzle: can God make a rock so heavy He can’t lift it? It works in all ten languages we tried, and also lights up for the unstoppable force (the phrase alone is enough) and the spear-and-shield story. It seems to fire while a puzzle could be coming and…","st":"tested","h":"../notes/L37_N11511/","i":8,"c":7,"m":[["rock paradox, Portuguese",64.5],["Bob's unliftable rock",55.8],["Can God create a",51.2],["immovable object",45.0],["spear and shield",25.8],["Can Mary create a",1.52],["almighty dollar",0.79],["not a strategy",-31.2]]},{"id":"L37_N134","k":"d","L":37,"N":134,"t":"A neuron for 'it would not be wrong to say' hedges","l":"This neuron responds to a small family of stock hedges that deny a claim is an error or an overreach: 'it would not be wrong to say', 'would it be fair to say', 'go out on a limb', 'not going too far'. It peaks on the predicate just before the claim begins. Form matters more…","st":"tested","i":5,"c":8,"m":[["overreach idiom",27.9],["question form",26.2],["core litotes",26.1],["present tense",26.0],["no negation",23.5],["person subject",22.4],["noun subject",18.6],["modal, no error",3.9],["control",-0.1],["modal praise",-6.0]]},{"id":"L37_N14860","k":"d","L":37,"N":14860,"t":"A neuron for great stores of written learning, led by Alexandria","l":"This neuron responds to a great store of written learning, and the Library of Alexandria is its strongest case. What matters is what the store holds, not how big it is: a library of scrolls fires strongly, while a treasury of the same size barely registers. It fires hardest on…","st":"tested","i":7,"c":8,"m":[["book collection",36.2],["library",30.1],["store of learning",27.0],["school",21.8],["archive",21.2],["nonsense",6.2],["treasury",5.3],["granary",4.6],["museum",4.4],["harbor",2.9]]},{"id":"L37_N15895","k":"d","L":37,"N":15895,"t":"Windows-developer context neuron with a memorized DC-suburb negative pole","l":"This neuron responds to a code or shell token when earlier text has placed it in the Microsoft developer world. Examples are the '3' of python3 after a PowerShell or cmd.exe setup, and 'public' inside a C# I-interface. Words that name the ecosystem itself, such as Azure or…","st":"tested","i":4,"c":3,"m":[["Azure",12.6],["nonsense",12.1],["Visual Studio",11.9],["PowerShell",11.3],["bash",11.1],["cmd.exe",11.0],["notebook",11.0],["Ubuntu",9.9],["zsh/macOS",8.8]]},{"id":"L37_N16634","k":"d","L":37,"N":16634,"t":"Too many to count: swarms that darken the sky","l":"This neuron fires when text implies that a swarm of small moving creatures was too numerous to count. Typical cases are locusts or pigeons darkening the sky, covering the fields, or flying 'so thick you could' knock them down. The same rhetoric about sheep, crowds or smoke…","st":"tested","i":8,"c":7,"m":[["locusts",34.5],["pigeons",26.4],["arrows",24.6],["starlings",19.5],["nonsense",18.2],["crowd",17.6],["smoke",14.6],["sheep",13.8],["snow",12.8]]},{"id":"L37_N17414","k":"d","L":37,"N":17414,"t":"Visiting the sick, and Little Red Riding Hood's name","l":"This neuron responds to people reaching out to the sick: a parish, volunteers or family visiting, phoning or bringing communion. It fires most when the outreach clause closes, not on the word 'sick'. The same neuron has a separate lexical trigger, the name Little Red Riding Hood…","st":"tested","i":9,"c":4,"m":[["clinicians",67.5],["kin",58.5],["parish",58.0],["LRRH name",55.2],["bikur cholim",52.5],["chaplaincy",52.0],["nonsense",51.0],["airway staff",50.0],["tradesman",48.0]]},{"id":"L37_N19448","k":"d","L":37,"N":19448,"t":"A folk luck ritual, just before its promised effect","l":"This neuron fires when text describes someone performing a familiar folk custom for luck, love, healing or a return trip, such as rubbing a statue, knocking on wood, clipping a love lock or tossing a coin in a fountain. It peaks where the text is about to say what the custom…","st":"tested","i":9,"c":8,"m":[["love lock",77.0],["rub statue",44.0],["coin toss",43.5],["rag tree",41.2],["prohibition",41.0],["mundane rub",35.0],["no ritual",24.9],["prayer",18.5],["hung charm",14.7],["fiction magic",6.9]]},{"id":"L37_N20391","k":"d","L":37,"N":20391,"t":"Lot's wife and the moment magic turns people to stone","l":"This neuron tracks one idea: magic turning a person into lifeless matter. Its prototype is Lot's wife, and her name alone drives it to about 80. Elsewhere it fires at the moment a wizard, curse or gaze is about to make someone stone, salt or gold, before the substance is named.…","st":"tested","i":9,"c":7,"m":[["wife of Lot",74.0],["Lot's wife",69.5],["Medusa victim",61.0],["nonsense",51.0],["Lot alone",48.0],["Orpheus look-back",47.8],["wizard curse",47.0],["other Bible wife",45.5],["mundane agent",44.0],["modern Lot",44.0]]},{"id":"L37_N22082","k":"d","L":37,"N":22082,"t":"A neuron for Apollo 11's first steps on the Moon","l":"This neuron tracks one historical moment: humans first walking on the Moon. It fires hardest on the words of Armstrong's 'one small step' quote, in any language and even in parodies. It also fires on the date July 20, 1969 and on 'first person to walk on the Moon', often before…","st":"tested","i":7,"c":6,"m":[["Apollo 11 crew",49.2],["future Moon",48.8],["Soviet space",47.0],["Apollo 11 (generic)",39.2],["later moonwalkers",23.9],["future Mars",21.1],["nonsense",15.1],["no-cue control",10.0],["Everest",9.1],["South Pole",4.8]]},{"id":"L37_N23362","k":"d","L":37,"N":23362,"t":"Turing-test scenes light up the slot before the judge","l":"This neuron tracks whether a Turing-test scene is in play and fires on open function words inside it, usually just before the judge or the candidate is named. Two cues each switch it on: Turing vocabulary (the name, the Loebner Prize, interrogators), or a machine trying to pass…","st":"tested","i":7,"c":8,"m":[["name + scenario",71.5],["Turing test",70.5],["no machine",67.5],["parlour game",63.5],["Loebner Prize",57.8],["machine vs human",49.5],["CAPTCHA",25.8],["blind tasting",15.2],["quiz control",15.0],["nonsense",10.6]]},{"id":"L37_N23570","k":"d","L":37,"N":23570,"t":"Snowflakes and 'no two alike': a lexeme plus a proverb","l":"This neuron responds to snowflakes and to the saying that no two are alike. The snow-crystal word fires by itself. The 'no two X are alike' construction fires for many subjects and fires most for the ones the proverb expects, such as leaves and fingerprints. When a snowflake…","st":"tested","i":7,"c":6,"m":[["snowflakes",40.5],["fingerprints",34.2],["snow crystals",33.8],["handmade goods",30.4],["zebras",28.5],["leaves",28.4],["nonsense",28.0],["paperclips",24.4],["pebbles",23.6],["cornflakes",21.4]]},{"id":"L37_N3297","k":"d","L":37,"N":3297,"t":"A neuron that expects a door-to-door round before it happens","l":"This neuron anticipates a round of stops. When context says someone is about to go door to door, house to house, office to office, or 'make the rounds' of talk shows, it fires on the connector ('from', 'the', 'по') before the venue appears. A strong purpose, such as searching…","st":"tested","i":7,"c":8,"m":[["search",31.1],["Halloween",21.4],["job hunt",19.4],["book tour",19.4],["duration only",19.1],["canvass",14.6],["terrain",7.0],["nonsense",3.0],["commute",0.9],["math fold",-19.5]]},{"id":"L37_N7051","k":"d","L":37,"N":7051,"t":"A clause about a dreamed-of total fix, signed by whether it is real","l":"This neuron asks one question of a clause: is it about a single fix that would make something hard easy or complete, and is that fix unreal? Denied shortcuts, wished-for miracle pills, and legendary cure-alls all fire it strongly, in any language, peaking on the word that…","st":"tested","i":8,"c":8,"m":[["mythic fix",12.9],["universal cure",9.5],["get-rich wish",8.5],["coveted goal",2.7],["nonsense control",-1.0],["mundane control",-2.0],["hard skill",-2.7],["skill advice",-2.9],["no-fix control",-2.9],["sober task",-4.0]]},{"id":"L38_N10532","k":"d","L":38,"N":10532,"t":"Told trickster and comic tales, with a small dip on scripture citations","l":"This neuron tracks told stories with a cunning or comic hero. It is strongly positive for Coyote, Anansi, Birbal and other tricksters and collections of funny anecdotes, and it fires again on later tokens, even the year in a citation. The names alone are silent, as in zoology…","st":"tested","i":7,"c":6,"m":[["Birbal",26.8],["Anansi",24.4],["Coyote tale",21.9],["funny anecdotes",15.2],["fairy tale",3.2],["Samson riddle",3.2],["Judges scripture",1.5],["secular history",1.1],["zoological coyote",0.4],["Rashomon",-7.0]]},{"id":"L38_N10591","k":"d","L":38,"N":10591,"t":"A rating-widget neuron that marks the vote count and predicts 'average'","l":"This neuron tracks the rater-count field of a rating summary: the 'how many people voted' slot beside a score. It peaks right after the count, as in '(6 votes, average: …', and still fires when the widget is empty or the number is bare. Counts behind other pooled figures, such…","st":"tested","i":6,"c":7,"m":[["votes (core)",70.0],["ratings",54.5],["sample count",34.8],["widget count",32.0],["engagement",23.9],["bare count",20.1],["nonsense noun",17.6],["sports tally",15.4],["word, no count",6.9],["non-count",6.3]]},{"id":"L38_N10635","k":"d","L":38,"N":10635,"t":"A neuron for letters moving out of order, not fixed slots","l":"This neuron responds when written units are moved out of order. It fires most for letters shuffled inside a word, as in the Cambridge 'first and last letter' meme (up to about +108) and Pig Latin rules. Whole words moved within a sentence give a weaker response, and non-text…","st":"tested","i":6,"c":4,"m":[["letters",72.5],["words",33.5],["sounds",21.4],["chapters",17.2],["digits",15.7],["crossword",14.1],["recipe",5.8],["songs",5.2],["guests",5.0],["grid slots",4.5]]},{"id":"L38_N14686","k":"d","L":38,"N":14686,"t":"Staying up to watch the sky versus ordinary night activities","l":"This neuron tracks one signed axis of night. It fires most strongly on the time slot that sets a viewing window, such as 'watch until [4]' in a meteor-shower guide. It fires more weakly for other sky-watching and for people sitting up all night. It goes negative when night is…","st":"tested","i":6,"c":7,"m":[["comet",20.2],["Perseids",15.2],["sitting up",15.1],["shooting stars",10.6],["paddleboarding",-2.7],["bakery control",-3.0],["lunar eclipse",-4.8],["harvest (neutral)",-8.0],["floodlit golf",-19.4],["night flowers",-30.4]]},{"id":"L38_N15011","k":"d","L":38,"N":15011,"t":"Beings who mistake their small world for all of reality","l":"This neuron fires on one scene: beings shut inside a smaller world (a cave, a flat surface, a bowl, a valley) who take it for all of reality while a larger world lies beyond them. Plato's chained prisoners and Flatland's 2D beings drive it hardest, near +50. Naming Plato or…","st":"tested","i":10,"c":8,"m":[["2D beings",37.8],["cave prisoners",37.2],["ant on balloon",24.0],["simulation",21.6],["goldfish bowl",17.9],["nonsense",12.2],["valley tribe",8.4],["office control",6.3],["coffee growers",-9.8],["silk farmers",-11.0]]},{"id":"L38_N15589","k":"d","L":38,"N":15589,"t":"Hoist by your own petard: a neuron for when your own side turns on you","l":"Fires when a sentence reveals that someone's trouble came from their own side: a leader killed by his own creation, stars turning on their creators, a judo throw using the attacker's weight. It peaks on words like \"own\" or \"very\". An outsider as culprit keeps it quiet; a…","st":"tested","h":"../notes/L38_N15589/","i":8,"c":7,"m":[["own creation",47.8],["the very people",44.0],["opponent's momentum",35.5],["own guards kill",29.1],["own guards protect",9.69],["killed by rebels",1.57],["own company",0.15]]},{"id":"L38_N18074","k":"d","L":38,"N":18074,"t":"A neuron for heat dosed on the vulnerable: Icarus and brooder chicks","l":"This neuron tracks how much heat a heat-sensitive thing is getting. Its clearest case is Icarus: wax wings flying too close to the sun, in any language, peaking near +60. Next come brooder instructions for chicks and piglets, and damageable things kept at a distance from a heat…","st":"tested","i":8,"c":5,"m":[["Icarus",57.2],["wax wings",50.2],["brooder chicks",36.2],["meltable",22.9],["moth idiom",13.2],["piglets",9.6],["chair (object)",9.1],["plain heat",5.5],["no cue",4.1],["geology (neg)",-4.5]]},{"id":"L38_N22132","k":"d","L":38,"N":22132,"t":"Known multi-return library call, halved once the code has unpacked it","l":"This neuron recognises the name of a library function the model knows returns several values at once: matplotlib's subplots, Go's os.Open, Lua's gsub, Rust's channel, Django's get_or_create, React hooks. It needs the library identity, since a user-defined function bound the same…","st":"tested","i":7,"c":6,"m":[["Rust channel (core)",18.6],["Go os.Open (core)",12.1],["subplots (core)",12.0],["Lua gsub (core)",9.1],["React hook",3.0],["generic multi-return",2.5],["Go single-return",1.0],["Python IO control",0.1],["user-defined control",0.1],["randint (negative)",-4.3]]},{"id":"L38_N23045","k":"d","L":38,"N":23045,"t":"A detector for deliberate alliteration, scaled by evidence of intent","l":"This neuron fires on words that continue a run of one initial letter, but only in proportion to how deliberate the run looks. Explicit frames like 'Tongue twister:' or 'Alliteration exercise:', dense marked vocabulary and long runs push it higher, while casual common-word echoes…","st":"tested","i":7,"c":8,"m":[["alliteration exercise",28.9],["tongue twister",26.8],["letter-game request",25.8],["poem",8.2],["bare",6.3],["bureaucratic frame",5.6],["nonsense frame",5.3],["headline",4.5],["diary",2.5],["grocery note",0.7]]},{"id":"L38_N23753","k":"d","L":38,"N":23753,"t":"A neuron for the letters i and e, named as a pair","l":"This neuron responds when a text names the letters i and e as a pair, and it peaks on the second letter. The textbook case is 'i before e except after c'. It starts rising at 'i before', before any e appears, and it helps the model write the rule's next step. A few memorized…","st":"tested","i":6,"c":5,"m":[["capital I",41.2],["i (core)",39.5],["letter i",38.8],["letter between",28.2],["no letter",21.2],["vowel a",20.5],["consonant b",18.6],["other vowel",15.6],["rule mimic",12.8]]},{"id":"L38_N25114","k":"d","L":38,"N":25114,"t":"An expectation neuron for Elijah's still small voice at Horeb","l":"This neuron expects the Horeb scene of 1 Kings 19 to come next. It fires most on the word just before the next element is due: the wind, earthquake, fire, or still small voice. Elijah alone raises it only a little. What makes it fire strongly is a frame that makes the scene's…","st":"tested","i":8,"c":8,"m":[["Tishbite",116],["Elijah",106],["Elijah + flight",104],["Moses",33.0],["other prophet",26.5],["Samuel",23.5],["conquered king",10.5],["secular person",3.3],["control",1.9]]},{"id":"L38_N650","k":"d","L":38,"N":650,"t":"One final surprise: a neuron that waits for the twist","l":"The strong pole is negative and anticipatory: it fires when a story, maze, exam or prosecutor has held something back to spring at the end, dipping already on 'one final' and deepest on the noun naming the reversal, surprise, twist or trap, in any language, good or bad. The…","st":"tested","i":8,"c":6,"m":[["door (near zero)",-2.9],["nonsense control",-5.8],["reward (pleasant)",-8.8],["trap",-13.3],["clue (hidden to find)",-14.6],["lesson (control)",-14.8],["puzzle (weaker)",-17.9],["Spanish surprise",-19.0],["twist",-21.2],["surprise (core)",-24.0]]},{"id":"L38_N7111","k":"d","L":38,"N":7111,"t":"Horseshoe crab blood and other living sources of biomedical materials","l":"This neuron fires when a biomedical material comes from a living source, and it responds most strongly to horseshoe crab blood used for endotoxin testing. It peaks on the slot where that use should appear next (about +40 at 'bled so that labs can'). Lab enzymes from microbes…","st":"tested","i":7,"c":5,"m":[["Limulus cells",28.1],["crab blood",25.5],["invented eel",21.1],["antiserum",19.6],["Taq source",19.4],["venom",15.2],["heparin pig",10.1],["recombinant",8.0],["factory",6.4],["invention",6.4]]},{"id":"L38_N946","k":"d","L":38,"N":946,"t":"The slide deck as a made thing, not the talk given","l":"This unit fires when text treats a slide deck as an object someone makes: PowerPoint, Keynote, Chinese PPT, 'Slide 3', or 'it' once 'it' means the deck. It is about meaning, not spelling: microscope slides and ship decks stay silent, while posters and photo slideshows get a…","st":"tested","i":6,"c":6,"m":[["Keynote",25.0],["presentation",21.8],["PowerPoint",19.2],["slides",18.9],["slideshow",11.4],["poster",8.1],["report",6.3],["spreadsheet",2.8],["recording",1.8],["microscope",-0.6]]},{"id":"L39_N13814","k":"d","L":39,"N":13814,"t":"Count versus list end: positive just short, negative at match","l":"This neuron reads a number in a count slot ('we have run N', 'among the N', a size argument in code) and compares it with where the list just above ended. It goes positive when the count lands a few steps short of the end, as with 'run 5' after Test 1 to 8. It goes strongly…","st":"tested","i":10,"c":7,"m":[["5 (core)",14.5],["3",12.5],["6",12.1],["7 (end-1)",7.7],["1 (far below)",-0.8],["12 (far above)",-2.6],["no number",-3.3],["9 (end+1)",-9.3],["8 (= end)",-24.8]]},{"id":"L39_N14696","k":"d","L":39,"N":14696,"t":"A Tower of Babel neuron that peaks before the story's next word","l":"This neuron tracks the model's belief that the Tower of Babel episode is being told. It fires across such passages and peaks at the slot just before the episode's own words, such as what they will build or which language is confounded. It needs neither the name Babel nor the…","st":"tested","i":7,"c":4,"m":[["post-flood",57.8],["Shinar, one language",53.2],["Babel",49.8],["Babylon",48.0],["Zeus transplant",45.0],["nonsense",41.5],["Exodus bricks",31.6],["beach control",21.4],["Dubai modern",17.1],["Popol Vuh",15.4]]},{"id":"L39_N20735","k":"d","L":39,"N":20735,"t":"A neuron that completes limericks and eLife digest openers","l":"This neuron tracks two memorized genre templates. The first is the limerick: the word itself and its stock opening, 'There once was a man from'. The second is the eLife-style science digest, as in 'Now, in eLife, X and colleagues report'. It fires on the slot whose next token…","st":"tested","i":6,"c":4,"m":[["limerick label",86.0],["other template",86.0],["generic poem",76.5],["limerick request",76.0],["prose control",75.0],["nonsense",74.5],["song lyric",70.5],["other form",70.0],["nursery rhyme",68.0],["hymn",67.0]]},{"id":"L39_N20764","k":"d","L":39,"N":20764,"t":"Seen first, heard later: a flash-to-bang schema neuron","l":"This neuron tracks a single schema: one event, usually lightning, reaches an observer twice, first as light and then as sound. It fires most on the sight verb that opens a flash-to-bang count, often before the event is named. It also fires, more weakly, on physics analogs such…","st":"tested","i":8,"c":2,"m":[["when you see",17.8],["senses swapped",16.8],["flash-to-bang count",15.2],["thunder came after",12.8],["light-first physics",12.5],["saw it, no timing",6.6],["storm words",2.1],["struck by",1.2],["payment network",-0.1],["hockey team",-0.2]]},{"id":"L39_N24284","k":"d","L":39,"N":24284,"t":"Italian beach lists versus seeking to the end of log files","l":"This neuron tracks text that is listing bathing beaches. How strongly it fires depends on where the beaches are: Italian seaside and lake beaches reach about 15, and every other place tested scores lower. Its peak lands on the first list number after the header. A separate…","st":"tested","i":5,"c":6,"m":[["Lake Como",16.8],["Puglia",15.2],["Sardinia",15.0],["Tuscany",13.9],["Italy (word)",11.9],["Swiss lake",10.8],["nonsense",9.3],["Crete",8.9],["Florida",6.8],["Portugal",4.5]]},{"id":"L39_N6628","k":"d","L":39,"N":6628,"t":"Mid-formula moment detector, graded by how second-order the average is","l":"This neuron fires while text is mid-way through writing the average of a random quantity, and fires harder the more second-order that average is. A bare E[X] barely registers, E[X^2] or E[XY] is moderate, and a variance or uncertainty formula combining two means peaks near +20.…","st":"tested","i":6,"c":6,"m":[["bracket Var",18.2],["bracket mean",15.6],["inner product",14.4],["Var formula",13.6],["E[XY]",11.9],["E[X^2]",8.8],["named statistic",4.8],["E[X]",4.5],["probability control",4.1],["prose moment",3.6]]},{"id":"L39_N7661","k":"d","L":39,"N":7661,"t":"Concrete thing in a slot that expects a process noun next","l":"This neuron fires on a word for a concrete physical thing, such as glacier, cotton, droplet or branch, when the sentence frame says a process or property noun about it comes next: 'the rate of glacier', 'to minimize branch'. Both are needed. Abstract nouns in the same slot earn…","st":"tested","i":5,"c":6,"m":[["cotton (material)",10.1],["glacier (core)",9.4],["horse (organism)",8.4],["table (everyday)",7.4],["determiner added",5.3],["soil (borderline)",4.8],["high school (sign flip)",4.5],["poverty (abstract)",3.9],["money (abstract)",2.2],["inflation (is the process)",-0.1]]},{"id":"L39_N9127","k":"d","L":39,"N":9127,"t":"A seam detector for theme words hidden inside pun coinages","l":"This neuron fires where a topic word surfaces inside another word or name, as in 'fish' inside Sofishticated or 'purr' inside Sir Purrcival. Visible respelling is the strongest cue, and a pun or naming frame adds to it. Ablating the neuron at the seam cuts the model's chance of…","st":"tested","i":9,"c":7,"m":[["off-theme pun",25.8],["hyphen splice",23.0],["respelled pun",21.8],["homophone swap",15.9],["plain theme name",6.8],["non-joke blend",4.9],["correct spelling",2.6],["math phrase",2.5],["nonsense",1.7],["neutral name",0.2]]},{"id":"L3_N21701","k":"d","L":3,"N":21701,"t":"A memorized-name neuron: large negatives on Beavis and Santorum suffix pieces","l":"This neuron recognizes a few rare names that the tokenizer splits into pieces, chiefly Beavis, Leavis and Santorum. It fires strongly negative on the second piece (down to about -60), and that signal helps the model predict what usually follows, such as ' and' after Beavis. The…","st":"tested","i":4,"c":7,"m":[["Grist leak",1.0],["Objectives leak",0.3],["Travis (control)",0.2],["whole token",0.0],["Sanctorum",-16.2],["Reavis (weak)",-17.4],["-vorous",-30.0],["Leavis",-37.2],["Santorum",-44.2],["Beavis (core)",-56.0]]},{"id":"L3_N2287","k":"d","L":3,"N":2287,"t":"A next-token gate keyed by net/净 and 某, signed by the partner","l":"L3/N2287 fires on the token directly after a key word. The key decides whether the neuron fires at all, and the current token decides the sign. After net or 净, content words like capital push it strongly positive, peaking near +43 on 净资本. After Chinese 某, person and classifier…","st":"tested","i":4,"c":9,"m":[["净 (CJK)",39.5],["net",34.2],["NET caps",34.0],["off-topic net",22.2],["net-ish stem",8.7],["某 key",5.0],["brand stem",0.7],["indefinite control",-0.5],["finance control",-0.6],["no key",-0.6]]},{"id":"L40_N16598","k":"d","L":40,"N":16598,"t":"A neuron for forwarded puzzles, read before the punchline","l":"This neuron fires while a text is setting up a recreational puzzle or curiosity of the kind passed around on puzzle pages and in email forwards: the border plane-crash riddle, the 'Take 1000, add 40' sum, fake 2=1 proofs, crazy-English lists, the railroad-gauge chain. It peaks…","st":"tested","i":7,"c":7,"m":[["academic write-up",47.0],["news report",44.0],["forwarded riddle",40.8],["chain-letter frame",39.8],["nonsense control",39.2],["trick question",38.0],["novel teaser",36.2],["puzzle page",36.0],["chat joke",32.0]]},{"id":"L40_N1809","k":"d","L":40,"N":1809,"t":"Position-one sink neuron whose positive side is leakage after repeated tokens","l":"This neuron tracks position, not meaning. Its only large event is a plunge on the second token of every sequence, deepest after a question word, produced when a positive gate and a large negative up-projection both read the layer-6 attention-sink units at position 0. Everywhere…","st":"tested","i":3,"c":9,"m":[["four 7s (run)",2.3],["two 7s (run)",1.7],["symbol",-8.0],["chat role",-8.1],["name",-10.7],["task verb",-12.0],["word fragment",-12.2],["3 EOT + What",-12.2],["question word",-14.7],["same token",-16.2]]},{"id":"L40_N19644","k":"d","L":40,"N":19644,"t":"Things that don’t add up: the TARDIS, and 1 + 2 + 3 + … = −1/12","l":"This neuron’s two favourite things are the TARDIS, bigger on the inside than the outside, and the claim that adding up every whole number gives −1/12. Both are totals that can’t be right. It fires on the setup, before the punchline, in nine languages, and set high it turns a…","st":"tested","h":"../notes/L40_N19644/","i":10,"c":4,"m":[["TARDIS, harvest",45.8],["TARDIS in Chinese",45.0],["all numbers, negative",33.5],["a==1 && a==2",26.6],["x is 1 and 2",17.6],["Hilbert's hotel",11.9],["sum 1 to 10",0.89],["Monty Hall",0.73]]},{"id":"L40_N21323","k":"d","L":40,"N":21323,"t":"Predicts the opening word of a legendary birth portent","l":"This neuron predicts that a birth portent is about to be stated: the dream, omen or seer that marks a child as a future king or Buddha. Its value builds as three cues accumulate: a mother or king from Greek, Persian, Indian or Chinese legend, a pregnancy or coming birth, and a…","st":"tested","i":8,"c":7,"m":[["Liu Bang's mother",35.5],["Olympias",31.0],["Maya (Buddha)",20.9],["Mandane",20.8],["Victoria",8.0],["Hecuba",5.8],["Mary (biblical)",4.4],["invented queen",3.3],["modern mother",3.0]]},{"id":"L40_N22150","k":"d","L":40,"N":22150,"t":"Completes the stock epithet slot for an established epic figure","l":"This neuron fires inside the fixed epithet formula for an epic god or hero, once the figure has already been established. It peaks on the open slot and on the stock attribute: 'Hera of the white arms' reaches about +80, while an uncanonical attribute like 'tired' stays near…","st":"tested","i":8,"c":7,"m":[["Athene",70.0],["Hera (core)",46.2],["Sanskrit epic",31.4],["'the goddess'",12.8],["Norse/OE epic",11.2],["modern name 2",6.5],["Virgil",2.3],["modern name",1.5],["sacred title",1.0],["no figure",-0.1]]},{"id":"L40_N6657","k":"d","L":40,"N":6657,"t":"A signed reader of index lists: subject headings up, dumps down","l":"This neuron reads lists of index terms. A list frame, such as a 'Descriptors:' or 'Keywords:' field or a category footer, opens the gate. The items then set the sign. Scholarly subject headings, especially ERIC education terms and DTIC engineering descriptors, drive it strongly…","st":"tested","i":4,"c":5,"m":[["DTIC caps",18.5],["history of X",16.9],["economics",16.8],["education",15.7],["virtues",13.7],["JoVE header",11.4],["nonsense",10.5],["scientist names",9.7],["news buzzwords",7.0],["India names",6.8]]},{"id":"L41_N10595","k":"d","L":41,"N":10595,"t":"Double-space reader: layout gap positive, deleted value negative","l":"This neuron fires on the token right after a stray double space, and its sign reports how the model interprets that gap. After a list item, column header, label or chat-turn marker, the gap is layout and the neuron goes positive. Inside a prose sentence where a number or…","st":"tested","i":6,"c":7,"m":[["chat turn",2.0],["bare list",0.6],["value present",-1.6],["column header",-2.2],["measure list",-3.0],["measure label",-3.2],["stripped citation",-4.7],["open prose slot",-8.5],["stripped measure",-9.9]]},{"id":"L41_N11223","k":"d","L":41,"N":11223,"t":"Reference knowledge versus a selling firm as a fact's source","l":"This neuron fires when a text reports a fact, and its sign tracks who stands behind that fact. It is positive when a fact is presented as reference knowledge, as in 'According to', 'Research shows' or 'As of June 2008' in an official register. It turns strongly negative when the…","st":"tested","i":9,"c":8,"m":[["gov agency",2.3],["university",-3.1],["bare name",-3.3],["polling firm",-4.2],["charity",-4.2],["research firm",-5.5],["sock company",-9.4],["software co.",-9.6],["payments firm",-11.5],["shoe retailer",-12.8]]},{"id":"L41_N13377","k":"d","L":41,"N":13377,"t":"A neuron for letting the bread dough rise","l":"This neuron fires on the verb for yeast dough rising, treated as a stage of making bread: 'let it rise', 'the dough rose', 'laissez lever'. It is strongest once the text has established the dough. Tense and mood do not matter, but other kinds of rising (sun, prices) do nothing.…","st":"tested","i":5,"c":7,"m":[["yeast dough",36.5],["pizza dough",34.0],["bare",32.5],["bloated belly",32.2],["cake batter",31.5],["kneaded clay",29.0],["prices",28.8],["pasta dough",24.6],["sun",22.2],["soup",19.1]]},{"id":"L41_N18391","k":"d","L":41,"N":18391,"t":"A 'that member doesn't exist' doubt signal in Qwen3-32B","l":"L41/N18391 fires when text credits an official offering with a plausible member the model doubts exists, such as Array.prototype.median() or collections.OrderedSet. It tracks the model's doubt, not the truth: recent real APIs like itertools.batched can fire as high as weaker…","st":"tested","i":10,"c":6,"m":[["fake: median",34.2],["fake: chunk",29.0],["fake: shuffle",27.4],["fake: sum",27.0],["fake, near .at",21.1],["nonsense",7.5],["real, recent",5.6],["real: includes",1.3],["real: sort",1.2]]},{"id":"L41_N23668","k":"d","L":41,"N":23668,"t":"Of all the days: a neuron for luck that picks its moment","l":"Fires on the words that single out the exact moment luck picked: the day we moved in, the moment I light the barbecue, the one day we visited. Loudest when the teller saw it coming (\"of course\", \"naturally\", \"knowing my luck\"). Real text uses this almost only for bad luck, but…","st":"tested","h":"../notes/L41_N23668/","i":8,"c":8,"m":[["wear white, spill",57.8],["free entry, the one day we visited",51.5],["of course, move-in day",27.5],["of course, sun at the beach",14.8],["rain at the beach, told flat",7.81],["plain bad luck",1.65],["Murphy's law itself",0.62]]},{"id":"L41_N9557","k":"d","L":41,"N":9557,"t":"Recognizes famous lateral-thinking puzzles from their memorized template wording","l":"L41/N9557 recognizes memorized lateral-thinking puzzles from their stock wording. It peaks on the noun that completes the template ('hotel' after 'pushes his car to a', 'water' after 'asks the bartender for a glass of') before any twist is given. Swapping one template word…","st":"tested","i":8,"c":7,"m":[["drives (swap)",55.5],["car→hotel",53.5],["bar→water",53.2],["novel scenario",27.4],["whiskey (swap)",23.4],["plain control",19.9],["liar logic",11.6],["number puzzle",6.0],["word riddle",4.0],["ListNode code",-9.4]]},{"id":"L42_N10733","k":"d","L":42,"N":10733,"t":"A neuron for topic-grounded puns, opposed to invented comic premises","l":"This neuron fires when a pun lands: a word from the document's own topic that suddenly carries a second, figurative sense, such as 'no resistance' from electricians or 'really scary' treats in a dentist's Halloween tips. The same word in a topic that doesn't license the double…","st":"tested","i":9,"c":8,"m":[["engineers",16.6],["electricians",14.8],["physics",14.6],["framed joke",12.9],["satire",7.8],["florists",2.4],["library",1.6],["nonsense",-2.3],["silly name",-2.5]]},{"id":"L42_N15611","k":"d","L":42,"N":15611,"t":"Haiku first-line break versus the driver in a scaling law","l":"This neuron marks the moment a haiku's first line ends. Once a context has announced a haiku, tanka or senryu, it fires on the break token after the roughly five-syllable opening line, peaking near +57. Other short verse forms, overlong first lines and later line breaks get only…","st":"tested","i":9,"c":8,"m":[["haiku",47.0],["senryu",44.0],["tanka",43.5],["5-7-5",35.8],["cinquain",23.0],["limerick",10.3],["short-line prose",8.1],["generic poem",1.8],["no-verse control",-0.4]]},{"id":"L42_N21766","k":"d","L":42,"N":21766,"t":"Compass-leg walks and where the traveller ends up","l":"This neuron tracks travel written as compass-direction legs with real distance units, such as 'one mile south, one mile east, one mile north'. It grows when the direction changes and peaks where the text is about to say where the traveller ends up. The model seems to read that…","st":"tested","i":9,"c":6,"m":[["pole destination",58.8],["steps",53.2],["single leg",53.0],["miles, 3 legs",51.8],["km, 3 legs",50.2],["Mars rover",50.2],["grid cm",8.8],["no directions",1.5],["landmark loop",0.8]]},{"id":"L42_N9250","k":"d","L":42,"N":9250,"t":"As much land as an oxhide can cover, and other fine print in legends","l":"Fires when a story spells out the terms of a legendary deal, strongest on asking for only as much land as a hide, cloak or day's walk can cover. Other fairy-tale conditions get about half; loans and leases stay near zero. Set high, it sometimes makes the model invent such deals.","st":"tested","h":"../notes/L42_N9250/","i":8,"c":5,"m":[["Dido (Spanish)",63.5],["Dido's cow hide",59.5],["Vamana's paces",44.5],["Oracle loophole",35.8],["Orpheus' condition",32.0],["Hatful of beans",24.9],["Beans as price",2.38],["Bail condition",0.53]]},{"id":"L43_N10546","k":"d","L":43,"N":10546,"t":"Fun-or-boring judgments: positive for chores and dentists, negative for leisure","l":"This neuron responds to judgments about whether something is fun, boring or frightening, and the object being judged sets its sign. For dreaded chores it goes strongly positive: taxes, forms, MLA citations, statistics homework, the dentist. This holds whether the verdict is…","st":"tested","i":9,"c":7,"m":[["statistics",16.6],["taxes",15.8],["MLA",12.6],["dentist",10.8],["chemistry",7.8],["laundry",5.2],["technical control",1.6],["no object",-2.5],["opera",-3.9],["golf",-7.4]]},{"id":"L43_N13927","k":"d","L":43,"N":13927,"t":"A detector for affirmed, weighted need, whatever the topic","l":"This neuron's main signal is negative. It fires on the word that asserts something is genuinely needed or warranted: 'there is a great need', 'is there a need', 'long overdue', 'the attention it deserves', 'no one needs more'. What matters is the construction, not the topic: an…","st":"tested","i":5,"c":2,"m":[["important",0.3],["optional want",0.3],["shortage",-0.4],["bare verb",-0.7],["negated need",-1.5],["great need",-4.8],["overdue",-5.2],["deserves",-5.4],["pressing need",-8.2],["trivial subject",-8.9]]},{"id":"L43_N2937","k":"d","L":43,"N":2937,"t":"A neuron for the amount being given away","l":"This neuron tracks one situation: part of what someone holds passes to another party, whether given, lent, traded, or put into a shared pool. It fires at the word slot where the transferred amount goes. Strong firing requires the situation to be set up, for example by repeated…","st":"tested","i":6,"c":7,"m":[["sold",16.6],["lent",15.6],["gave (core)",13.7],["pool",9.1],["pop (negative)",5.8],["no recipient",4.4],["spent",3.1],["lost",3.1],["kept (control)",-5.6],["took from pile",-5.8]]},{"id":"L44_N12472","k":"d","L":44,"N":12472,"t":"The last piece of a Japanese researcher's given name","l":"Fires on the final subword of a romanized Japanese given name when it opens a reference to a present-day professional, usually a researcher or author. Japanese surnames, places and companies in the same slot stay near zero. The frame matters as much as the name: 'Dr.' or a…","st":"tested","i":7,"c":7,"m":[["Japanese given",21.2],["invented Japanese",20.8],["Korean given",15.9],["Chinese given",14.6],["Italian given",11.7],["Japanese loanword",8.2],["familiar Western",3.4],["Japanese surname",1.0],["Japanese place",0.7],["Arabic given",-3.9]]},{"id":"L44_N12501","k":"d","L":44,"N":12501,"t":"Introduce new coordinates (positive) or single out one variable (negative)","l":"In math solutions this neuron tracks how the solver moves variables. Writing a fresh joint description of the whole object, like 'Introduce u = x + y, v = x − y' or cutting a cube into eight identical tetrahedra, drives it positive. Singling one letter out as the argument and…","st":"tested","i":7,"c":6,"m":[["transformation",4.6],["new coordinates u,v",4.2],["congruent pieces",1.3],["no-math control",0.3],["symmetrize",-2.6],["solve for y",-10.4],["joint unknowns",-11.6],["one letter, no '='",-12.9],["fix x, iterate y",-17.4],["quadratic in x",-18.0]]},{"id":"L44_N20524","k":"d","L":44,"N":20524,"t":"Fires when the model decodes a ciphertext word, especially shifted 'the'","l":"This neuron fires on the last token of a word hidden in English-shaped ciphertext when the model can tell what the word is. The strongest trigger is a Caesar-shifted 'the' such as 'uif' or 'xli'. A made-up cipher word that has recurred also triggers it. The memorized shifts…","st":"tested","i":10,"c":8,"m":[["shift-22 'the'",66.5],["shift-1 'the'",60.8],["shift-4 'the'",26.4],["ROT13 (memorized)",9.8],["Caesar-3 (memorized)",9.7],["shift-1 'you'",6.1],["shift-1 'and'",4.1],["random letters",3.7],["leetspeak 'the'",0.2],["plaintext 'the'",0.2]]},{"id":"L44_N23322","k":"d","L":44,"N":23322,"t":"A phrase-closure gate whose sign is set by the passage","l":"This unit's gate opens on the token that closes a word, phrase or sentence in text presented for reading or solving, and its up projection assigns a sign that belongs to the particular passage rather than to a topic. Positives are rare fixed phrases inside exam problems. The…","st":"tested","i":2,"c":2,"m":[["geoboard say (+)",3.5],["satellite axis (+)",-0.9],["plain prose",-1.2],["geoboard for example",-1.4],["farm control",-1.5],["array photographs",-1.9],["coin tosses (+)",-2.5],["nonsense",-3.2],["array values (-)",-13.9],["orchard (-)",-18.6]]},{"id":"L44_N23645","k":"d","L":44,"N":23645,"t":"A second, unrelated person with the same surname","l":"L44/N23645 fires when a newly introduced person reuses a surname already worn by someone else earlier in the text, with nothing setting up the match: a second Skrip on a roster, two Shevchenkos in a reference list. A kinship cue halves the response. A partner introduced with a…","st":"tested","i":10,"c":8,"m":[["uncued copy",32.2],["title+surname",6.2],["cued sister",6.1],["same person",4.6],["non-person copy",2.4],["no name",-0.5],["near-miss surname",-0.6],["fresh name",-0.7],["spouse, fresh name",-2.3]]},{"id":"L44_N4053","k":"d","L":44,"N":4053,"t":"Sending items to sides by value, not picking the middle","l":"This neuron fires when many items are being sent to two containers by a test on their value. It peaks on the destination word, as in 'everything larger goes in the right sub-array', which reaches about +38. Quicksort is its densest source, but sorting worksheets, MapReduce keys…","st":"tested","i":7,"c":6,"m":[["< pivot",19.9],["by color",16.2],["< 50",14.3],["even value",13.6],["coin flip",12.8],["nonsense",10.2],["even index",5.5],["first half",5.1],["before mid",4.9],["middle index",-6.2]]},{"id":"L44_N8869","k":"d","L":44,"N":8869,"t":"A neuron that marks where a pasted code snippet closes","l":"This neuron tracks nesting in code and markup pasted into a post. It fires hardest at the first moment a snippet's outermost element is clearly closing: the dedent back to the outer column, or the outer closing tag name such as '<\/plugins>' (about +29, against +9 for the inner…","st":"tested","i":6,"c":7,"m":[["root close",19.4],["second close",17.5],["inner close",6.0],["dedent",6.0],["scope open",6.0],["prose control",0.2],["brace close",-4.0],["under wrapper",-7.9],["html open",-12.1]]},{"id":"L45_N12561","k":"d","L":45,"N":12561,"t":"Throw it, hide it, rub it: hand-performed luck charms","l":"This neuron tracks a specific kind of folk custom: a person does a small act with their hands to an object, and the act is supposed to work by magic rather than by physics. Examples are throwing a tooth onto the roof, hiding a coin in the pudding, or rubbing a statue's nose. It…","st":"tested","i":6,"c":6,"m":[["tooth on roof",23.8],["leap fire",20.1],["secular chore",17.8],["rub statue",17.2],["coin in pudding",15.2],["hung garlic",10.4],["lanterns",5.5],["sweeping taboo",5.2],["pray to spirits",5.2],["nonsense",3.9]]},{"id":"L45_N15555","k":"d","L":45,"N":15555,"t":"Predicts 'wish I'd found this sooner' after praise; negative predicts proposals","l":"This neuron predicts one continuation: after someone praises a helpful thing, a wish that they had had it earlier. It fires on 'I wish I' or 'Where was this...' before the earliness words arrive, reaching about +60 when praise, a wish and a time anchor combine. Too-late…","st":"tested","i":8,"c":7,"m":[["praised course",30.2],["praised app",29.4],["too-late fix",11.4],["person met",7.8],["broken app",6.2],["envy",4.2],["bad luck",3.0],["neutral",2.2],["proposal",-22.8]]},{"id":"L45_N16022","k":"d","L":45,"N":16022,"t":"A bereavement axis: the one who lost someone versus the loss itself","l":"In a passage where someone has just been killed or carried off, this neuron reads the role of the current token. Words that designate the bereaved person, a father, a king, the fiancée of the victim, fire strongly negative, and the clause about them stays negative regardless of…","st":"tested","i":10,"c":6,"m":[["the captor",-1.2],["no-person control",-1.2],["official, not kin",-1.3],["non-kin control",-3.2],["kin relation",-5.2],["loss named",-5.4],["bereaved by title",-6.6],["bereaved sibling",-7.6],["feeling named",-9.0],["bereaved kin",-13.0]]},{"id":"L45_N16800","k":"d","L":45,"N":16800,"t":"A neuron that notices when a word has two readings","l":"This neuron notices when a word has just been read two ways: a name returning as an ordinary word, a topic word reused in an idiom, or a name whose origin shows in the text. It fires on the next token, often the start of 'no pun intended' or 'hence the name'. If a naming frame…","st":"tested","i":8,"c":8,"m":[["root echo",16.5],["eponym-like",15.1],["compound echo",13.8],["no name",6.3],["opaque name",-9.1],["technical term",-10.6],["synonym compound",-11.0],["separate synonym",-15.1],["nonsense name",-15.1]]},{"id":"L45_N19244","k":"d","L":45,"N":19244,"t":"A tongue-twister neuron: one word reused in a new role","l":"This neuron in Qwen3-32B fires when a word returns within a few tokens in a new, readable grammatical role, as in 'I thought a thought' or 'a canner who can can cans'. Memorized twisters such as Peter Piper or the woodchuck also trigger it, in several languages. Plain…","st":"tested","i":9,"c":7,"m":[["modal→verb",56.8],["verb→noun",49.0],["memorized",48.2],["digit reuse",33.2],["alliteration",27.1],["name→noun",23.2],["spread apart",14.8],["verbatim",2.8],["no reuse",1.4],["earnest echo",0.8]]},{"id":"L45_N9295","k":"d","L":45,"N":9295,"t":"Signed word-onset fragment unit: prose fragments positive, named-member slots negative","l":"L45/N9295 responds to the first piece of a word or name that is split across several tokens, and the frame sets the sign. Inside running text whose context already points to one completion ('Wordsworth. Bi', 'B9090 West'), it fires positive. When the fragment opens the next…","st":"tested","i":4,"c":7,"m":[["road code",7.0],["prose fragment",5.8],["member slot, Th",2.0],["whole word",1.2],["name, finished",1.2],["member slot, Hall",0.5],["function word",0.5],["member slot, 2nd",-4.0],["member slot, 1st",-4.5]]},{"id":"L46_N12856","k":"d","L":46,"N":12856,"t":"A neuron for pieces taken in order to read as one","l":"L46/N12856 fires when a text sets up separate pieces to be taken in order and read or joined as one whole. Acrostics are the clearest case ('the initial letters of each line, taken in order'). Connect-the-dots, plotting points to join, reassembled packets and album tracks taken…","st":"tested","i":8,"c":5,"m":[["acrostic letters",17.6],["gallery paintings",11.6],["album songs",11.0],["plotted points",10.6],["numbered dots",10.3],["wall stones",6.0],["house rooms",3.2],["nonsense",2.0],["train cars",-2.0],["polyhedron faces",-4.9]]},{"id":"L46_N15614","k":"d","L":46,"N":15614,"t":"Digit 8: positive in serial codes, negative as a quantity","l":"This neuron watches the digit 8, and more weakly 3. Context sets its sign. When the 8 is part of a code, such as an ISSN, ISBN, DOI or journal volume, it fires strongly. When the same 8 is a quantity (an operand, a data value, a year), it turns negative. Changing only the label…","st":"tested","i":6,"c":6,"m":[["DOI",22.2],["ISSN",21.5],["eISSN",21.2],["PMID",6.6],["phone",1.3],["ISBN",0.7],["total",-0.7],["year",-0.8],["order no.",-2.0],["dividend",-6.2]]},{"id":"L46_N16280","k":"d","L":46,"N":16280,"t":"A direction bit for two-way relations: counterpart stated versus roles swapped","l":"This neuron keeps track of which way a two-way relation is being stated. When text names one side of a pair and then supplies its counterpart, especially inside a rule or definition (too high, so too low; the domain of f is the range of f inverse), it fires strongly positive…","st":"tested","i":7,"c":8,"m":[["verbatim repeat",13.6],["counterpart (core)",12.9],["connective after rule",12.9],["third-item control",12.9],["vice versa",12.9],["roles swapped",12.9],["same value repeated",6.4],["off-axis control",2.8],["loose partner value",1.1]]},{"id":"L46_N2246","k":"d","L":46,"N":2246,"t":"A signed tracker: first write in a code arm versus closing it","l":"This neuron tracks the state of the innermost if/else or loop arm in code. It turns positive the moment the first short action in that arm finishes, such as `count += 1` or a push. It turns negative on the continue, break, raise, or return that closes the arm right after such an…","st":"tested","i":5,"c":6,"m":[["dict write",9.2],["second stmt",8.9],["write+break",8.9],["write+raise",8.9],["py append",7.2],["first write",4.2],["bare return",-0.1],["comment control",-0.4],["bare continue",-2.2],["write+continue",-11.6]]},{"id":"L46_N24358","k":"d","L":46,"N":24358,"t":"A signed dial for who speaks next: asking negative, answering positive","l":"This neuron tracks who is about to take the floor in a two-party exchange, and which way. It plunges when the text invites, permits, or forbids the other party to ask: letter closings like 'Please let me know if' reach about -20, and the role tag opening a reply is also…","st":"tested","i":6,"c":5,"m":[["answer directive",1.0],["no-cue letter",0.3],["nonsense",0.3],["non-asking verb",-0.6],["prohibition",-2.1],["contact invite",-5.0],["core sign-off",-5.2],["ask verb",-6.2],["comment invite",-7.2]]},{"id":"L46_N24715","k":"d","L":46,"N":24715,"t":"Problem tokens: what to compute (+) versus which candidate to choose (−)","l":"Inside formal problems, this neuron reads the role of each token. It fires positive where the thing to be worked out is named or written: 'which quantity is labelled X', the expression to evaluate, or the values in answer choices. It fires negative on scaffolding that sets up a…","st":"tested","i":5,"c":6,"m":[["value",2.3],["variable",1.3],["quantity (core)",0.9],["property",0.9],["vague noun",0.6],["colour",0.5],["physical quantity",0.3],["nonsense",-1.6],["option (choice)",-4.1],["given options",-5.1]]},{"id":"L46_N331","k":"d","L":46,"N":331,"t":"Comparison connective: identity of particulars up, ranking of kinds down","l":"This neuron fires on the word that opens the standard of a comparison ('as', 'than', 'like') and its sign says which kind of comparison it is. Sameness or a stated multiple between particular concrete things drives it positive ('the same widgets as'). Ranking a generic kind on a…","st":"tested","i":5,"c":6,"m":[["same N as, particular",10.2],["more than, particular",6.3],["N times the age",5.9],["same, kind",2.6],["less than, particular",-0.5],["no comparison",-0.5],["less than, concrete kind",-2.8],["'like' examples",-7.1],["shortfall, abstract kind",-8.5],["less than, kind",-10.1]]},{"id":"L47_N10376","k":"d","L":47,"N":10376,"t":"A 'nothing left over' neuron with a school-list side job","l":"This neuron has two poles. Its large, causal pole is negative: it fires on the constraint clause of equal-sharing word problems ('identical sets… with no… left over'), and zeroing it there knocks 'none' from the model's likely next words. Its weaker positive pole fires on list…","st":"tested","i":6,"c":5,"m":[["total count",-15.1],["subtraction",-15.1],["rate problem",-15.1],["no math",-18.8],["bare GCD",-19.6],["school list",-27.0],["equal tiling",-36.8],["identical sets",-56.5],["equal groups",-65.0]]},{"id":"L47_N21265","k":"d","L":47,"N":21265,"t":"The future that is always just around the corner","l":"This neuron knows the joke that fusion is always thirty years away. It fires on the 'when' of a forecast that keeps being made and never comes true. Said once, a promise leaves it flat; promised every year for a decade, it jumps to around +35. Set high, it makes the model talk…","st":"tested","h":"../notes/L47_N21265/","i":9,"c":7,"m":[["fusion, sixty years",81.0],["self-driving, a decade",75.5],["Linux desktop",57.0],["subway, every year",46.5],["contractor, four months",30.1],["garage 'tomorrow'",14.4],["subway, said once",0.5],["opened on schedule",0.26]]},{"id":"L47_N22800","k":"d","L":47,"N":22800,"t":"A neuron for turning from individual cases to the whole","l":"This neuron marks the moment a text stops handling things one at a time and turns to the option that covers them all: a general formula after worked values, a subscription after a single purchase, a change of character after separate apologies. It peaks on the verb or determiner…","st":"tested","i":6,"c":6,"m":[["aphorism halo",17.1],["worked values",13.4],["one check",5.6],["one vs all",5.4],["placeholder",1.1],["sequential",0.3],["no cue",0.2],["select one",-0.2],["separate apologies",-0.3],["no contrast",-0.3]]},{"id":"L47_N6182","k":"d","L":47,"N":6182,"t":"Before the noun: one place, or something said of every cell","l":"This neuron decides, one token early, how a sentence is about to locate things. It swings negative on the word that names or quantifies over a family of equal cells: groups of, each row, equivalence classes, GROUP BY. It swings positive on the article or preposition that opens a…","st":"tested","i":6,"c":6,"m":[["one box (core)",3.5],["the row",2.8],["one place",2.6],["vague one",2.6],["all cells",2.6],["each cell",2.6],["every cell",2.6],["one line",-0.3],["no place",-0.3],["packed units",-1.3]]},{"id":"L47_N8344","k":"d","L":47,"N":8344,"t":"The 'but' before providence: invisible hand and Genesis 50:20","l":"This neuron marks a hinge. It fires just after a text states an agent's narrow intention in the wording of two famous formulas: Smith's trader who 'intends only his own gain' and Joseph's 'you meant evil against me'. The peak comes right before a hidden guiding agency, the…","st":"tested","i":9,"c":6,"m":[["Genesis core",65.0],["Smith quote",43.2],["paraphrase",40.8],["Smith core",38.0],["providence prose",21.2],["Genesis, good",20.0],["emergent order",5.9],["control",0.9],["nonsense",0.7],["direct divine",-7.3]]},{"id":"L48_N11494","k":"d","L":48,"N":11494,"t":"A count word that opens a coined letter label","l":"This neuron fires on a count word, such as 3, three, 三 or tres, when the model expects it to open a coined label like \"the 4 Rs\" or 三严三实. It is strongest after a list whose items share an initial letter, and it pushes the next token toward that letter rather than the default…","st":"tested","i":9,"c":8,"m":[["novel P list",35.2],["novel G list",34.8],["memorized Rs",28.8],["2-item list",21.0],["empty",17.4],["4-item R list",16.4],["plain list",15.7],["no list",10.1],["nonsense list",7.2]]},{"id":"L48_N12050","k":"d","L":48,"N":12050,"t":"Defining a twin term (+) versus recalling a conversion factor (−)","l":"This neuron fires when technical writing sets a term apart from its near-twin: the real versus the nominal interest rate, pore velocity, absolute pressure, or truth versus satisfaction. It peaks at the slot where the definition begins, not on the qualifier. Its negative pole is…","st":"tested","i":4,"c":4,"m":[["pore velocity",9.9],["real rate",4.9],["real exch.",4.2],["real wage",3.6],["real GDP",3.0],["abs. pressure",2.7],["nonsense",2.2],["net weight",1.6],["mass (textbook)",0.6],["conversion",-4.5]]},{"id":"L48_N16046","k":"d","L":48,"N":16046,"t":"A neuron that gets ready to state Occam's razor","l":"This neuron fires at the point where the parsimony maxim is about to be stated, as in 'the maxim that | entities should not be multiplied', and it stays active while that wording is spelled out. A frame naming Occam's razor and the formula itself each suffice, and together they…","st":"tested","i":8,"c":8,"m":[["Occam's razor",81.0],["lex parsimoniae",76.0],["least privilege",61.8],["minimal effort",48.5],["Hanlon's razor",41.0],["Murphy's law",28.8],["no principle",13.0],["Kant",10.7],["Newton (descriptive)",10.3],["club rule",9.9]]},{"id":"L48_N16774","k":"d","L":48,"N":16774,"t":"A neuron that fires where a dream's contents begin","l":"This neuron tracks the contents of a sleep dream as they are being told. It stays quiet on the word 'dream' and peaks on the first token of what the dream contains, at about +40 in plain reports and up to about +100 in dream-symbol lists. Visions, daydreams, imagined scenes and…","st":"tested","i":7,"c":8,"m":[["nightmare",36.0],["sleep dream",34.2],["night vision",30.2],["daydream",24.5],["imagination",7.8],["while asleep",7.0],["photo scene",2.7],["nonsense",0.5],["mundane control",0.2],["waking control",-0.1]]},{"id":"L48_N18669","k":"d","L":48,"N":18669,"t":"Fires when one new item follows a plural group","l":"This neuron fires on the word that brings in one new item right after a plural group: 'she bought shirts and a', 'Pipes A and B … Pipe C'. Order and countability matter. Singular-first lists read about zero, and paired or bulk goods like shoes or nails read weakly. Word problems…","st":"tested","i":7,"c":8,"m":[["count plural",33.5],["explicit count",19.1],["pairs restored",17.2],["paired goods",14.7],["bulk-ish plural",14.2],["or control",10.5],["collective",5.1],["paired goods",2.8],["singular first",0.2]]},{"id":"L48_N20304","k":"d","L":48,"N":20304,"t":"A Chan/Zen dialogue detector that fires at the master's reply","l":"This neuron tracks how strongly the text reads as a Chan/Zen encounter dialogue. Named Chan masters, the 'A monk asked' formula and Chan questions like 'What is Buddha?' each push it up, and together they saturate it. It peaks at the quote where the master's reply begins, and…","st":"tested","i":8,"c":8,"m":[["Chan name",54.0],["Chan master",52.0],["Zen, unnamed",26.9],["Thai forest",16.1],["Indian Buddha",15.1],["lama",10.0],["Desert Father",8.4],["secular master",4.0],["nonsense",1.8],["plain name",1.1]]},{"id":"L48_N20479","k":"d","L":48,"N":20479,"t":"A neuron that expects boundary conditions after a differential equation","l":"This neuron fires when a differential equation has been set up as a problem whose answer will be pinned by side conditions: a boundary-value, eigenvalue or initial-value problem. It peaks on the slot where those conditions would start, such as the trailing '= 0', ' with' or a…","st":"tested","i":7,"c":7,"m":[["Green's fn",35.5],["equilibrium",27.1],["eigenproblem",21.5],["no-math control",17.8],["BVP",17.5],["constraint halo",16.1],["IVP",12.8],["plain ODE",-5.3],["general solution",-20.5]]},{"id":"L48_N24284","k":"d","L":48,"N":24284,"t":"Next case pending after a run of passes, versus failures","l":"This neuron tracks a series of parallel cases checked one at a time, such as ledger requests, test checks or repeated log lines. It fires on the token that finishes stating the next case, before its verdict. It is positive when the earlier cases passed and grows with the length…","st":"tested","i":7,"c":6,"m":[["mixed run",6.9],["3 passes",5.2],["1 pass",4.2],["2 passes",2.1],["uniform tag",2.1],["no verdicts",1.6],["3 failures",1.4],["no run",0.7],["unrelated prose",-0.2]]},{"id":"L48_N253","k":"d","L":48,"N":253,"t":"Integers to be counted or printed under a rule","l":"This neuron tracks integers treated as things to produce, whether said aloud in turn, printed in a loop or written in some notation, under a rule about which ones appear. It fires hardest at the point where that rule begins, as in FizzBuzz or the Sevens clap game. Skipped room…","st":"tested","i":7,"c":7,"m":[["count aloud",29.9],["letters",20.6],["months",14.8],["print 1..N",14.4],["counting apples",12.9],["cyclic turns",12.8],["room numbering",9.6],["summing",5.4],["word taboo",4.5],["digit slots",-8.0]]},{"id":"L48_N5650","k":"d","L":48,"N":5650,"t":"A '1→8' digit detector that fires inside phone numbers and IDs","l":"This neuron fires on a digit '8' that directly follows a '1', but only strongly when the string reads as a phone number or formal identifier such as an ISSN, ISBN or grant code. A frame word alone can switch it on or off. The same bigram in a year or an age gives a weak…","st":"tested","i":4,"c":5,"m":[["phone 1→8",91.0],["ISSN 1→8",69.5],["grant 1→8",57.5],["phone 3→8",36.2],["year 1→8",11.1],["order no. 1→8",7.5],["room 4→8",-1.0],["PMID 4→8",-37.5],["ISBN 2→8",-46.5]]},{"id":"L49_N11083","k":"d","L":49,"N":11083,"t":"Bare 'A boy' subjects versus person descriptions about to come","l":"This neuron tracks whether a person in the text is described or left bare. It fires positive on the token right after an undescribed stock character, most strongly 'A boy [is]' opening a word problem. It fires negative at the moment a description is about to be supplied: the 'a'…","st":"tested","i":5,"c":8,"m":[["bare boy (core)",17.5],["bare girl",9.9],["definite",8.9],["bare man",8.7],["one adjective",4.0],["object control",2.0],["generic person",-1.2],["weak noun",-1.2],["named",-1.9],["demographic adj",-6.3]]},{"id":"L49_N11562","k":"d","L":49,"N":11562,"t":"Values or declarations: a member-boundary detector inside record literals","l":"L49/N11562 fires at member boundaries inside brace or indent-delimited blocks and asks one question: do the members hold values or declare fields? It is positive throughout JSON, dicts, object literals and YAML, peaking on the whitespace before the next key or a nested closer.…","st":"tested","i":5,"c":8,"m":[["Vue options",15.5],["JS object",14.3],["JSON record",10.2],["dict opener",5.8],["function body (control)",0.3],["prose (control)",-1.3],["Python class",-1.7],["typed fields (halo)",-4.4],["CodingKeys enum",-5.6],["GraphQL type",-5.8]]},{"id":"L49_N11629","k":"d","L":49,"N":11629,"t":"A neuron that waits for the forbidden act in superstitions","l":"This neuron tracks inherited folk rules, superstitions above all. Once text starts passing on a taboo or omen, it fires on each token right before the forbidden act or omen is named ('it is unlucky to…'), then falls once the act appears. The more folk-like the source, the…","st":"tested","i":8,"c":6,"m":[["superstition",31.0],["folklore",26.8],["grandmother",24.2],["beekeeper",17.8],["scripture",12.5],["nonsense",9.2],["store manager",8.9],["math",8.1],["company policy",3.8],["copyright",-1.1]]},{"id":"L49_N14189","k":"d","L":49,"N":14189,"t":"Memorized openings lock in; a backdrop clause waiting for 'when' goes negative","l":"This neuron's strongest firings come from a short list of stock texts the model knows by heart, each at the one token where it becomes unmistakable: the comma after '1, 2, 3, 4', the 'interest' after 'percent annual'. Swap the names and rates and it still fires; break the…","st":"tested","i":6,"c":2,"m":[["set comma (core)",12.6],["annual interest (core)",11.5],["practising for his",11.0],["for her (weaker)",7.7],["fairy-tale opening",4.1],["action + when (control)",0.9],["fronted when (control)",0.8],["nonsense control",0.2],["plain word problem",-0.4],["backdrop state",-3.9]]},{"id":"L49_N21124","k":"d","L":49,"N":21124,"t":"One name axis: cited authorities fire negative, performers fire positive","l":"This neuron sorts names by one question: cite them or watch them? Completing the surname of a canonical thinker, historian, poet, or report-issuing body drives it strongly negative, and the sign carries onto the role noun that follows. Actors, celebrities, and scientists…","st":"tested","i":6,"c":3,"m":[["screen actor",3.4],["credited discoverer",2.5],["no-name control",2.2],["athlete control",-0.6],["mid-scale thinker",-1.6],["living novelist",-2.1],["living thinker",-2.2],["cited scientist",-3.7],["canonical poet",-4.8],["canonical historian",-7.3]]},{"id":"L49_N23305","k":"d","L":49,"N":23305,"t":"Copying the previous sibling's opener in code lists, not caching","l":"This neuron fires on a token that starts the next item in a list by copying the previous item's opener. Examples are the second key in { p1: 1, p2 }, the next argument in f(p1, p2), and this.cache in a repeated class. It needs a code or data frame, and the copy must sit on the…","st":"tested","i":5,"c":8,"m":[["key copy",27.6],["k copy",20.8],["arg copy",13.4],["cache copy",7.6],["new container",2.9],["newline break",2.4],["foo copy",2.2],["value slot",2.2],["no copy",2.2]]},{"id":"L49_N6103","k":"d","L":49,"N":6103,"t":"The pause after 'X is easy': granting a generic verdict","l":"This neuron fires on the punctuation or 'but' that closes a short, flat verdict about something generic, such as 'Designing a game is hard,' or 'Generosity is admirable;'. Such a verdict reads as a point to be granted before a contrast arrives. Specific subjects, hedges…","st":"tested","i":7,"c":6,"m":[["scope-limited",40.8],["bare activity",30.6],["virtue noun",28.6],["generic gerund",28.5],["specific activity",25.9],["nonsense",20.5],["hedged",8.2],["non-evaluative-ish",2.7],["specific object",0.3],["narrative",0.3]]},{"id":"L49_N6225","k":"d","L":49,"N":6225,"t":"A hedge token that asks who is about to judge this","l":"This neuron sits on the word just before a sentence says how something is regarded: 'is often', 'can be', or the comma closing 'Whenever I say that homeopathy works'. Its sign records who is judging. It goes mildly positive when the writer hands down or relays a general label…","st":"tested","i":8,"c":8,"m":[["nonsense control",-0.8],["anti-vaxxers say",-2.2],["I hear",-2.2],["I read",-3.5],["my critics say",-6.1],["doctors say",-10.3],["scientists say",-12.7],["I argue",-23.5],["I say",-29.0],["I write",-37.8]]},{"id":"L49_N7006","k":"d","L":49,"N":7006,"t":"Midweek-date predictor at the slot before a day-of-month number","l":"This neuron predicts a midweek date. On the token just before a day-of-month number, it fires strongly negative when context says the date falls on a Wednesday or Thursday. The cue can be a weekday word or a weekly journal's known publication day, such as NEJM, which reaches…","st":"tested","i":10,"c":5,"m":[["nonsense",-1.9],["BMJ (Sat)",-2.4],["Lancet (Sat)",-2.6],["Science (Fri)",-4.4],["magazine",-5.4],["Cell (Fri)",-9.6],["PNAS (Tue)",-12.6],["Nature (Thu)",-20.8],["JAMA (Wed)",-23.1],["NEJM (Thu)",-35.2]]},{"id":"L49_N784","k":"d","L":49,"N":784,"t":"Scholarly prose about to name itself, its authors, or sources","l":"This neuron fires in impersonal scholarly and expository prose at the moment just before the text names itself, its authors, or a cited source: 'This [study]', 'Here [we]', 'see [Figure]', a citation bracket. Register acts as the gate. A personal aside or chat frame shrinks the…","st":"tested","i":6,"c":8,"m":[["no frame",18.9],["abstract",14.2],["news",13.8],["slide outline",12.8],["textbook",12.2],["personal aside",10.7],["code string",9.4],["chat frame",9.3],["casual",6.8]]},{"id":"L49_N8993","k":"d","L":49,"N":8993,"t":"Flags the moment an earlier-named thing will explain something new","l":"This neuron fires in factual, explanatory writing at the moment a newly introduced thing has been described in a way that calls for a cause, and something named earlier is about to supply it. An example is a bell 'ringing continuously since 1840,' after a sentence about old…","st":"tested","i":6,"c":6,"m":[["old battery (core)",17.4],["template frame",12.6],["dry battery",11.8],["nonsense",5.3],["food analog",3.2],["poor-fit battery",1.1],["no context",-1.0],["fiction",-1.9],["topic, no explainer",-2.0]]},{"id":"L4_N14041","k":"d","L":4,"N":14041,"t":"Completes two-word 'X variable/number' compounds; the modifier sets the sign through the up-projection","l":"L4/N14041 fires on the second token of a two-word technical compound, and the first word decides sign and size. The center is 'random variable' in any surface form: +32 on ' variable' after 'random', +36 to +42 for 'Random Variable', 'randomvariable', 'random-variable' and…","st":"tested","i":4,"c":5,"q":[["p0_01",-0.553],["p0_1",-0.363],["p1",-0.085],["p5",-0.016],["p10",-0.016],["p25",-0.016],["p50",-0.016],["p75",-0.016],["p90",0.016],["p95",0.016],["p99",0.016],["p99_9",0.085],["p99_99",2.4]]},{"id":"L4_N24185","k":"d","L":4,"N":24185,"t":"A word-piece lookup: fires after 'Old', sinks on Burnley and otherworldly","l":"This neuron is a one-step spelling lookup, not a topic detector. It fires on the word-piece that immediately follows the piece 'Old' inside a single word, with a fixed strength per continuation: Oldways and Oldly are strongest, Oldbury and Oldridge are medium, and familiar…","st":"tested","i":3,"c":3,"m":[["Old|ways core",40.5],["Old|ies core",26.6],["Old|bury core",23.0],["Cold leak",8.4],["-ley control",-0.0],["spaced Old",-0.1],["Gold control",-0.1],["Old|ham silent",-1.0],["Burnley negative",-39.8],["otherworldly negative",-39.8]]},{"id":"L50_N10103","k":"d","L":50,"N":10103,"t":"Symbols subtracted read positive, named parties agreeing read negative","l":"This neuron reads two things placed in relation. When two paired symbols are subtracted or compared, it fires positive on the operator, about +10 at the '>' of |a-b| > 2, with no problem frame needed. When explicitly named parties act jointly, it fires negative on the verb and…","st":"tested","i":5,"c":6,"m":[["non-agent pair",-3.2],["states",-4.7],["named pair",-6.7],["pronoun",-6.7],["institutions",-8.2],["legal register",-8.2],["collective noun",-9.8],["three named",-10.0],["both",-13.2],["two players",-14.5]]},{"id":"L50_N1148","k":"d","L":50,"N":1148,"t":"Insurance coverage topic state, negative on every token, math problems faintly positive","l":"This neuron tracks whether the text is about being covered against loss by a contract. Any passage about an insurance policy, in advice, narrative or definition alike, pushes it strongly negative on nearly every token, reaching about -45 at the extreme. Warranties, protection…","st":"tested","i":5,"c":6,"m":[["sofa control",-7.3],["smoke detector",-9.1],["phone plan",-10.5],["AppleCare",-18.2],["helmet",-19.0],["warranty",-19.1],["dental plan",-22.1],["renter's insurance",-27.5],["makeup",-28.8],["travel insurance",-32.8]]},{"id":"L50_N12173","k":"d","L":50,"N":12173,"t":"One neuron separates owed service from neighbourly help","l":"This neuron reads text about one person helping or giving to another and sorts it by relationship. When peers help informally, neighbours lending tools or a village getting through winter, it swings strongly negative and its output promotes words like 'mutual' and 'share'. When…","st":"tested","i":7,"c":5,"m":[["hired labour",5.7],["Philemon",4.5],["cow trade",2.1],["tenancy",2.0],["institutional charity",1.6],["no-transfer control",0.7],["workplace sharing",-4.8],["scriptural giving",-5.2],["village mutual aid",-21.6],["peer lending",-23.8]]},{"id":"L50_N12788","k":"d","L":50,"N":12788,"t":"A 'beginning to' neuron that fires when the onset is still underway","l":"This neuron marks the 'to' of begin/start + to when the writer presents the onset as happening now, from their own vantage. Progressive marking roughly doubles its response, so 'she is beginning to' fires about eight times harder than a dated 'In 1998 she began to'. Transition…","st":"tested","i":6,"c":4,"m":[["progressive begin",46.0],["transition halo",42.8],["chronicle subject",40.2],["perfect begin",29.8],["simple past",21.6],["modal start",18.4],["rival aspect verb",5.2],["activity control",4.5],["reset start",-0.3],["web-notice being",-5.8]]},{"id":"L50_N14784","k":"d","L":50,"N":14784,"t":"A neuron that fires when the model takes a minimum","l":"This neuron turns on when the model is about to produce the smallest member of a set: a bounding box's left edge, a min column, xs.min(), or the 'smallest possible' answer to a contest problem. The operation is the cue, not the word: swapping min for max with all other tokens…","st":"tested","i":9,"c":7,"m":[["lowest",4.4],["minimum",0.9],["largest (mirror)",0.4],["no operation",0.4],["2nd smallest",0.3],["smallest (core)",-0.3],["middle",-0.5],["greatest",-0.5],["nonsense",-0.5],["least",-0.7]]},{"id":"L50_N17269","k":"d","L":50,"N":17269,"t":"A link-word neuron that restates quantities by name, and mirrors formulas negatively","l":"Deep in Qwen3-32B, this neuron fires on the small link word of a word-problem comparison, the 'of' in 'gave Sandy 221 of the balloons' or the 'the' in 'three times the amount', at the moment the model decides to name an earlier quantity again rather than say 'them'. Ablating it…","st":"tested","i":6,"c":6,"m":[["red marbles",21.8],["green balloons",18.9],["baseball cards",12.4],["nonsense noun",11.9],["chocolate cookies",11.5],["generic items",10.5],["measure control",8.9],["bare balloons",6.5],["money amount",5.9],["bare cookies",4.8]]},{"id":"L50_N17433","k":"d","L":50,"N":17433,"t":"Little did they know: a neuron for narrated past ignorance","l":"This neuron fires when past narration says someone did not yet know what was coming: 'Little did she know', 'had no idea', 'ne savait pas'. It peaks on the ignorance word itself, before the reveal, so the content that follows does not matter. A narrator's retrospective 'would'…","st":"tested","i":8,"c":7,"m":[["no idea",40.0],["little did",31.5],["was not to",28.8],["did not know",23.0],["present teaser",23.0],["not yet learned",16.4],["would",5.1],["forgot",2.4],["neutral",-0.4],["first step",-1.9]]},{"id":"L50_N20933","k":"d","L":50,"N":20933,"t":"Fires where a new blend name starts copying its first parent","l":"This neuron fires when a text is inventing a name for a mix of two things, on the token that starts copying one parent into that name: ' P'uggle, ' gaz'ellephant, 'Opp'enbarbie. The more the context expects a blend, the stronger it fires. A lone frame gives a small response, and…","st":"tested","i":10,"c":7,"m":[["whole-word parent",107],["cued dog blend",105],["animal blend onset",85.5],["non-parent word",51.2],["talk, no coinage",48.5],["onset, no merge",44.0],["real non-copy name",16.9],["fixed-word fragment",4.5],["math continuation",4.2]]},{"id":"L50_N2108","k":"d","L":50,"N":2108,"t":"Who is helping: institution as provider versus practitioner with their students","l":"This neuron tracks who is doing the helping in education and health text. It fires positive when an organization speaks as 'we' on a service page that a header has already framed, such as a homework-help site announcing its tutors. It fires negative when a specific teacher…","st":"tested","i":5,"c":6,"m":[["provider 'we'",6.0],["institution distributes",-3.0],["no cue (control)",-3.0],["parent uses",-4.8],["buyer, no role",-4.8],["plumber (control)",-15.4],["nurse uses",-15.6],["plural teachers",-22.5],["teachers (3rd person)",-23.0],["math teacher uses",-25.2]]},{"id":"L50_N3507","k":"d","L":50,"N":3507,"t":"Naming a math object versus filling a slot in its written form","l":"When a letter appears in a math statement, this neuron reads what job it does. It fires when the letter is the name of one object, strongest for a single capital letter right after a geometric figure, as in 'regular hexagon H', and in 'where q is prime' clauses. It swings…","st":"tested","i":7,"c":8,"m":[["regular hexagon",17.0],["graph (halo)",10.1],["circle",8.9],["point (falls off)",1.7],["ship (non-math)",1.0],["matrix (halo)",-1.0],["concrete sum (control)",-4.5],["a + bH (coefficient)",-8.8],["2^k H (product form)",-22.4],["p/H (fraction form)",-22.6]]},{"id":"L51_N13938","k":"d","L":51,"N":13938,"t":"Love as an ideal versus a source you draw comfort from","l":"A neuron about inner emotional life whose sign depends on how a bond is framed. Love, friendship, or the soul spoken of as a good in itself, in an aphorism or a philosopher's definition, drives it positive and it predicts a simile. A marriage, job, or religion narrated as…","st":"tested","i":6,"c":6,"m":[["love (core)",8.5],["friendship",2.9],["marriage",2.8],["family",1.1],["coffee (control)",-0.5],["no-bond control",-0.8],["music (halo)",-2.3],["job",-3.0],["faith",-8.4],["religion (neg core)",-13.7]]},{"id":"L51_N19918","k":"d","L":51,"N":19918,"t":"A signed reader of whether a human life is at stake","l":"This neuron rates what is at stake just before a text states the outcome. It fires strongly, up to about +50, on the slot before a hazard is said to kill or an action to save a human life. Statistics, near-death stories and reported executions give weaker positive values. At the…","st":"tested","i":7,"c":4,"m":[["cardiac arrest",23.1],["heat stroke",22.5],["cobra bite",14.7],["infection",13.2],["non-human",4.6],["figurative",2.4],["property",0.7],["no stake",-0.8],["non-fatal injury",-0.9]]},{"id":"L51_N20229","k":"d","L":51,"N":20229,"t":"Expects a right angle to be stated next; opposes upcoming side lengths","l":"This neuron predicts what kind of claim a geometry text will make next. It fires on the token just before a right angle is stated, but only when an earlier word such as rectangle, right triangle, perpendicular, or Thales has signalled it. Other angle slots get a weaker response.…","st":"tested","i":6,"c":8,"m":[["Thales cue",23.8],["square",22.0],["right triangle at B",20.0],["negated right",19.4],["obtuse",18.0],["rectangle",15.1],["arithmetic 90",12.8],["non-geometry 'right'",8.3],["equilateral",-10.0],["side lengths",-29.8]]},{"id":"L51_N23909","k":"d","L":51,"N":23909,"t":"A classifier reader: trade names fire, subset adjectives flip it negative","l":"This neuron reads the word that says what kind of physical thing is about to be named. It fires when that word is a trade, craft or clinical practice, such as embroidery before needles or orthopaedic before cement, and how-to prose roughly quadruples the response. Product grades…","st":"tested","i":5,"c":6,"m":[["embroidery (craft)",11.5],["landscaping (trade)",9.8],["orthopaedic (clinical)",8.4],["heavy-duty (grade)",7.1],["special (placeholder)",5.6],["steel (material)",2.7],["red (plain)",0.1],["some (no classifier)",-0.6],["equal (quantity)",-3.9],["human",-5.1]]},{"id":"L51_N25261","k":"d","L":51,"N":25261,"t":"The adverb slot after 'have' that predicts a participle","l":"This neuron fires on the word right after a form of 'have', such as 'also', 'both', 'not' or an inverted 'you', and on the contraction 've. At these spots it predicts that a past participle like 'been' or 'got' comes next. Removing it at 'Kale has also' roughly halves the…","st":"tested","i":6,"c":8,"m":[["had",70.5],["has (core)",61.8],["to have",53.5],["do-support",36.2],["habitual gap",34.0],["future",5.2],["modal",4.0],["copula",-1.2],["and (negative)",-2.7],["but (negative)",-3.5]]},{"id":"L51_N6851","k":"d","L":51,"N":6851,"t":"Separators inside word-split sentences; negative at math operand slots","l":"This neuron fires on the separator between two words of one sentence broken across lines, extra spaces or quoted list items. It fires when the word before is a content word or preposition, and stays quiet after function words. It weakens when the text repeats and can be copied.…","st":"tested","i":3,"c":6,"m":[["preposition",22.9],["content verb",17.6],["content verb",11.4],["article",7.3],["preposition",6.3],["title noun",5.2],["auxiliary",3.6],["nonsense",3.2],["operation word",-1.6],["pronoun",-3.3]]},{"id":"L51_N9673","k":"d","L":51,"N":9673,"t":"Opener of a licensed insertion, strongest before a pronoun slot","l":"L51/N9673 fires on the token that opens a short inserted chunk, such as ' (' in a catalog heading, ' [' in an edited quotation, or ' **' in bold-heavy chat, but only once the context has made that insertion conventional, either by genre or by one earlier instance. It fires two…","st":"tested","i":6,"c":6,"m":[["subject heading",17.6],["catalog list line",17.5],["one catalog line",16.9],["prior qualifier parens",5.7],["nonsense",-1.3],["chat answer",-3.0],["plain prose",-3.6],["year-list template",-3.7],["German gloss",-4.0],["one prior prose paren",-5.5]]},{"id":"L52_N11896","k":"d","L":52,"N":11896,"t":"Taking the fun out: qualities treated as removable substances","l":"This neuron fires when an abstract quality such as fun, boredom, stress, guesswork or life is treated like a substance that a verb pulls out of, or puts back into, an activity, as in 'this course takes the boredom out of accounting.' It peaks on the quality noun and helps…","st":"tested","i":8,"c":8,"m":[["take out",124],["suck out",98.0],["guesswork",84.5],["put in",77.5],["edge",71.0],["no particle",43.0],["amount change",32.0],["non-removal take",27.6],["static verb",10.5],["quantity verb",7.4]]},{"id":"L52_N12450","k":"d","L":52,"N":12450,"t":"Digit after an 8: identity sets the sign, frame sets the gain","l":"This neuron reads the digit that comes right after an 8 inside a number or identifier and reports which digit it is on a signed axis. A following 1 or 6 drives it strongly negative, deepest in phone numbers and standards codes like ISO/IEC 8652. A following 0, 5 or 7 is positive…","st":"tested","i":3,"c":3,"m":[["arXiv 8→0",29.4],["dot reset",1.1],["grant 8→0",0.3],["no digit",0.2],["pages 8→7",-1.2],["control 7→1",-3.0],["toll-free 8→0",-17.2],["phone 6→1",-28.8],["phone 8→1",-33.8],["phone 8→6",-40.5]]},{"id":"L52_N13995","k":"d","L":52,"N":13995,"t":"A game-guide axis: effects positive, player guidance and loot negative","l":"In video-game text this neuron reads whole clauses and picks a side. Text that tells the player what to do or how to get things, crafting, farming, unlock costs, tier lists, boss advice, pushes it hard negative, and that pole is the one that matters: it promotes acquisition…","st":"tested","i":5,"c":7,"m":[["grants shield",21.0],["deals damage",17.8],["restores HP",17.5],["no game control",7.9],["indicator value",-5.4],["grants XP",-8.1],["tier ranking",-20.6],["costs to unlock",-26.6],["crafted from",-44.5],["rare drop, farm",-59.2]]},{"id":"L52_N16622","k":"d","L":52,"N":16622,"t":"Reply-onset label token that signals more of the label is coming","l":"At the start of a reply, this neuron fires on the first token of a restated topic label, but only when that label is certain to continue: 'Nova' before Scotia, or the comma after '1' in a long chemical name. The bare, complete term does not fire, and the token that closes a…","st":"tested","i":5,"c":7,"m":[["two-piece name",22.6],["three-piece name",20.2],["nonsense",14.6],["two-word term",14.2],["single token + colon",10.6],["uncopied label",6.2],["chemical, comma after 1",4.9],["no label",2.5],["single token, bare",-3.2],["weak two-piece",-6.0]]},{"id":"L52_N21637","k":"d","L":52,"N":21637,"t":"A neuron that flips glyphs: upside-down and mirrored characters","l":"This neuron tracks a flip of a character's shape while it is happening. When the text announces a 180-degree rotation or a mirror image, it turns on and peaks at the slot just before the flipped glyph is written, as in 'the 9 becomes a| 6'. Zeroing it makes the model copy the…","st":"tested","i":10,"c":7,"m":[["upside down",112],["reversal",102],["rotated 180",100],["mirror",99.5],["cipher",73.0],["bare 6↔9",38.0],["nonsense",33.0],["neutral",28.2],["arithmetic",8.4],["card game",-41.8]]},{"id":"L52_N21766","k":"d","L":52,"N":21766,"t":"Junction neuron: next span re-uses given material or builds a new relation","l":"This neuron sits at clause junctions and predicts what kind of span comes next. It is positive at the word that opens a specification clause when that clause will re-mention things the passage already set up, such as 'the probability that' after a described random walk, or a…","st":"tested","i":5,"c":2,"m":[["ant on cube",3.8],["point in square",0.8],["fact control",0.8],["nonsense control",0.5],["marbles in bag",0.4],["coin sequence",-0.8],["number n",-1.1],["frog on pads",-2.7],["ε-N bound",-33.8]]},{"id":"L52_N23503","k":"d","L":52,"N":23503,"t":"A neuron that predicts the next words of Eden's Fall","l":"This neuron tracks the story of Adam and Eve's Fall. Within a telling of the Fall, from any tradition or stance, it fires on the word just before a canonical story element: 'tree of' before 'the knowledge of good and evil', the serpent's 'You will not surely die', or the colon…","st":"tested","i":8,"c":4,"m":[["serpent + Eve",80.0],["Genesis Fall",78.0],["Milton",66.0],["Revelation",54.0],["Adapa myth",53.0],["pop-culture Eden",48.8],["fruit botany",38.0],["fairy tale",35.2],["control",13.9],["wood use",8.0]]},{"id":"L52_N24188","k":"d","L":52,"N":24188,"t":"Predicts that a present-day oddity will be called a relic","l":"This neuron fires when the model is about to call something in the present a survival from the past. It peaks on the empty slot right before 'relic', 'vestige', 'holdover' or 'leftover', especially after the text has called the thing odd or useless. Sacred relics and leftover…","st":"tested","i":7,"c":4,"m":[["no history",54.2],["useless organ",45.2],["old tech",34.8],["cosmic remnant",31.1],["odd custom",21.1],["sacred relic",21.1],["still active",8.2],["literal leftover",5.7],["continued after end",4.3],["no-cue control",-1.5]]},{"id":"L52_N25095","k":"d","L":52,"N":25095,"t":"Distances and long-haul carriage, against 'the word came from'","l":"This neuron fires when text is about to say how far something travels between places. It peaks on the slot just before a city-to-city distance and on transport words like 'shipped' or 'made the trip' before the destination. Two memorized templates, 'need only look' and JEL code…","st":"tested","i":4,"c":3,"m":[["made the trip",25.0],["shipped",11.3],["came from",8.6],["transported",8.0],["carried",1.5],["priced (control)",1.5],["cleaned (control)",1.5],["made in",1.5],["etymology",-4.6],["term came from",-6.2]]},{"id":"L52_N3388","k":"d","L":52,"N":3388,"t":"Individuals about to form units, versus a named team's member count","l":"L52/N3388 tracks collectives of individuals along one signed axis. Its strong pole is negative and fires in the gap between an arranging verb and the unit word, when students, campers, wolves or stars are about to be put into groups, pairs or packs but the word has not landed…","st":"tested","i":5,"c":6,"m":[["data (abstract)",-1.3],["budget (mass)",-1.7],["territory (mass)",-5.8],["nonsense",-10.9],["named roster",-11.4],["wolves (animals)",-12.2],["stars (physical)",-16.2],["campers",-22.4],["participants",-24.8],["students (core)",-26.9]]},{"id":"L52_N343","k":"d","L":52,"N":343,"t":"After the celebration: a neuron for trouble arriving","l":"This neuron fires when a story announces that trouble has arrived. It peaks on the problem noun itself, such as 'trouble', 'problems' or 'doubts', and it fires hardest when a good situation came just before, as in 'celebrated for weeks. Yet trouble'. Static complaints and…","st":"tested","i":7,"c":6,"m":[["trouble (core)",23.1],["problems",21.4],["doubts",16.0],["disaster",13.1],["short-lived (halo)",11.3],["cracks",10.1],["joy (control)",2.1],["war (danger)",1.8],["fire (danger)",0.8],["no noun",0.1]]},{"id":"L52_N3653","k":"d","L":52,"N":3653,"t":"A factor-of-three neuron: positive for tripling, negative for stoichiometry","l":"This neuron tracks a factor of three, and its sign depends on what kind of three it is. It fires on the '=' and the first digit of a small, familiar tripling result (3 x 6 = 18; Hour 3: 27). At multiple-choice answer slots it can decide the next digit. Swap the factor for 2 or…","st":"tested","i":8,"c":6,"m":[["verbal ×3",16.9],["long operand",12.6],["3 x 6 (core)",10.8],["factor 6",2.4],["factor 4",0.9],["same 18, ×2",-1.0],["addition",-1.4],["no number",-2.5],["factor 9",-2.6],["chem coefficient",-33.5]]},{"id":"L52_N4551","k":"d","L":52,"N":4551,"t":"'By their own admission': a neuron for self-incriminating testimony","l":"This neuron fires strongly negative when an argument cites a party's own testimony against it, as in 'by the EPA's own…'. The peak lands on ' own', and the neuron pushes the next word toward 'admission'. The evidential frame matters more than who the party is. Zeroing the neuron…","st":"tested","i":8,"c":7,"m":[["spatial 'by'",-18.4],["Harvard, no frame",-26.9],["NASA, no frame",-30.1],["EPA, no frame",-37.8],["frame + EPA",-49.5],["frame + Exxon",-49.5],["frame + bakery",-50.5],["frame + nonsense",-51.2],["frame + opponents",-56.0]]},{"id":"L52_N8672","k":"d","L":52,"N":8672,"t":"Slot before a specified item: CJK fires, English quantity goes negative","l":"This neuron watches structured request and specification text (chat clarifications, input() prompts, field lists, parameter docs) at the slot just before the specified item is named. The sign depends on language. Korean and Japanese slots fire strongly positive, up to about +70…","st":"tested","i":7,"c":6,"m":[["Korean menu",47.0],["Japanese menu",36.8],["Chinese menu",7.2],["German menu",3.8],["English menu",3.5],["empty",-4.0],["nonsense",-4.7],["Russian menu",-5.0],["Spanish menu",-9.1],["English quantity",-13.2]]},{"id":"L52_N9853","k":"d","L":52,"N":9853,"t":"Global carbon cuts versus green-building energy performance","l":"This neuron is a signed axis in environmental text, and it fires on the slot just before a quantity. It goes positive when the quantity is greenhouse emissions of the climate system, fossil fuels, nations or the world, as in IPCC cut rates, lifecycle CO2 tables and climate…","st":"tested","i":6,"c":7,"m":[["world",5.0],["home",4.3],["nation/bloc",4.2],["fossil supply",4.0],["transport halo",4.0],["no-resource control",-2.1],["renewable site",-6.2],["city",-10.2],["building",-15.2],["green building",-46.5]]},{"id":"L52_N9958","k":"d","L":52,"N":9958,"t":"The moment before “were it not for”","l":"This neuron fires partway through a what-if, when a sentence has said what would have happened and is about to name the one real thing that stopped it: “would have gone hungry had it not been for the churches.” It goes quiet once the rescuer is named. Its negative side sits just…","st":"tested","h":"../notes/L52_N9958/","i":8,"c":8,"m":[["had it not",112],["were it not",103],["Russian если бы",89.0],["homeless today if",80.5],["dead today if",45.5],["millionaire today if",1.89],["thanks to levee",1.78],["word-problem what-if",-74.0]]},{"id":"L53_N20625","k":"d","L":53,"N":20625,"t":"The modal that sets up 'would make a sailor blush'","l":"This neuron fires on the word that opens a predicate about something the text has just made remarkable, as in 'a torrent of curses that [would]' or 'a vocabulary that [could]'. At that point the model is setting up a stock comparison such as 'make a sailor blush' or 'rival X'.…","st":"tested","i":8,"c":6,"m":[["torrent of curses",87.0],["shoe hoard",68.5],["PR blitz",49.2],["plain kitchen",44.5],["orgy of excess",37.2],["venomous letter",37.2],["nonsense",28.1],["plain letter",23.4],["long bridge",14.1]]},{"id":"L53_N20679","k":"d","L":53,"N":20679,"t":"Draw-from collections versus wholes that contain themselves","l":"This neuron reads a containment statement through the kind of holder. A shared gate opens whenever something 'contains' or 'holds' something, and the up projection sets the sign. Finite collections you draw from or test against, such as urns, decks, subsets, committees and…","st":"tested","i":7,"c":7,"m":[["urn (core)",11.6],["drawer",7.2],["deck",5.6],["jar (neutral)",-1.0],["box",-6.8],["theory",-10.8],["apple",-13.6],["seed",-15.4],["nonsense control",-16.0],["acorn (deep)",-31.6]]},{"id":"L53_N22779","k":"d","L":53,"N":22779,"t":"The 'none other than' neuron: firing just before a reveal","l":"This neuron marks the moment a text is about to reveal who someone really is. It peaks on the word 'none' in 'was none other than…' and fires more weakly on the words leading up to it. Its direct output favours emphatic-identity words in other languages, such as Chinese 竟是 and…","st":"tested","i":6,"c":7,"m":[["non-fiction",57.2],["hidden caller",55.8],["nonsense",52.8],["secret admirer",48.8],["culprit reveal",44.5],["notable author",43.2],["none-the-wiser idiom",28.8],["mundane copula",27.2],["numeric none",25.8]]},{"id":"L53_N277","k":"d","L":53,"N":277,"t":"Places opening to the public, versus products sold on a market","l":"This neuron tracks the moment a text says a place, such as a bridge, tunnel, rail line, cave or museum, was opened to traffic or to the public. It fires on ' opened' or ' to' and helps write the 'to the public/traffic' that follows. Ablating it at ' opened' cuts P(' to') by up…","st":"tested","i":6,"c":6,"m":[["bridge",21.5],["cave",20.5],["museum",8.9],["rail line",7.0],["nonsense",5.9],["album",5.7],["smartphone",5.0],["console",4.9],["bakery",-3.2],["TV show",-11.7]]},{"id":"L53_N4978","k":"d","L":53,"N":4978,"t":"Oral and sexual routes of infection, plus 'men who have'","l":"This neuron fires on the word that names how an infection travels through the body, mainly ' oral', plus anal and anogenital, in HPV, STI and infectious-disease text. A separate input lights it on ' have' in 'men who have sex with men'. Other organ adjectives give a weaker…","st":"tested","i":4,"c":4,"m":[["eye (organ)",22.8],["oral (core)",22.4],["oral-genital",22.4],["anogenital",21.8],["colorectal (organ)",21.5],["anal",13.0],["respiratory route",11.3],["nonsense",4.6],["non-body control",1.8],["no route word",1.1]]},{"id":"L53_N5496","k":"d","L":53,"N":5496,"t":"Stock-template object slots versus attributive 'central' awaiting its noun","l":"This neuron has two unrelated poles that share one gate; the up projection decides the sign. It turns positive on the function word just before a genre's stock object, as in word-problem openers, author disclosures, 'Build your' calls to action and '# Set the' in GUI setup code.…","st":"tested","i":4,"c":6,"m":[["word problem",22.4],["built his",17.9],["disclosure",13.9],["career CTA",6.9],["weaker noun",5.1],["pygame comment",4.7],["made his",-1.4],["yoga 'your'",-9.7],["peripheral",-13.9],["Central",-65.0]]},{"id":"L53_N555","k":"d","L":53,"N":555,"t":"The 'synthetic' neuron: fires negative on the word, promotes what follows","l":"This neuron detects the word synthetic. It swings negative on synthetic and synthesis in any spelling or language, anticipates on the bare onset Syn-, and dips mildly on textiles, polymers, and syn- lookalikes like Cynthia. At synthetic its output suppresses repeating the…","st":"tested","i":5,"c":8,"m":[["positive pole",19.8],["control",-4.5],["artificial",-5.3],["syn-, not synth",-11.5],["textile halo",-12.4],["synthesis",-67.0],["bare Syn-",-74.0],["Chinese 合成",-98.0],["synthetic (core)",-103],["misspelt",-106]]},{"id":"L53_N8961","k":"d","L":53,"N":8961,"t":"Own verdict versus others' construal: a sign-coded labeling neuron","l":"L53/N8961 fires just before a label is attached to something, and its sign says whose label it will be. It goes positive when a speaker is about to state flatly what a contested thing really is, usually a deflating verdict such as 'free will is just an illusion.' It goes…","st":"tested","i":7,"c":6,"m":[["skeptic verdict",14.8],["determinist",13.9],["hedged say",9.1],["non-deflating",9.0],["narrator verdict",7.6],["nonsense",6.5],["others' construal",2.8],["rule",1.1],["no frame",0.3]]},{"id":"L54_N11963","k":"d","L":54,"N":11963,"t":"A surname-onset neuron for ' Ay', damped by Jenny and Whitney","l":"This neuron reads the first piece of a surname. It fires strongly on ' Ay' when that token starts a person's last name, after a first name, an initial or a title, regardless of whether Ayers, Ayala or Aykroyd follows. The same Ay in places or ordinary words fires about half as…","st":"tested","i":3,"c":5,"m":[["initial + Ay",171],["first name + Ay",157],["word Ay",99.0],["Jenny + Ay",89.0],["place Ay",65.0],["Ak onset",31.8],["Ey onset",26.0],["other onset",3.5],["no name",-0.7],["Jenny + U",-22.6]]},{"id":"L54_N11965","k":"d","L":54,"N":11965,"t":"A neuron for hitching a ride on someone else's transport","l":"This neuron fires at the moment a text says something with no transport of its own gets carried by a bigger mover. Examples include a seed on a dog, a parasite in a mosquito, a satellite piggybacking on a rocket, and a stranded commuter getting a lift. It peaks on 'hitch' or the…","st":"tested","i":8,"c":5,"m":[["payload hitches",197],["seed hitches",163],["ride request",79.0],["stranded commuter",68.5],["bandwagon",64.5],["plain carriage",15.8],["own transport",10.1],["capacity count",4.8],["Ferris wheel",3.2],["no passenger",0.6]]},{"id":"L54_N14003","k":"d","L":54,"N":14003,"t":"Cooperators versus defectors, set against evolution as survival and descent","l":"This neuron tracks one axis inside evolution text. On its positive side are named strategy types competing in a population: cooperators and defectors, hawks and doves, mutants trying to invade. The Prisoner's Dilemma is its peak, especially right before a payoff number. Counted…","st":"tested","i":6,"c":6,"m":[["cooperators",108],["tit-for-tat",101],["resistant bacteria",78.0],["defectors",74.0],["hawks",59.8],["dark moths",58.5],["red allele",51.0],["survive/reproduce",50.5],["smartphones",50.2],["nonsense",38.2]]},{"id":"L54_N14698","k":"d","L":54,"N":14698,"t":"About to name a concrete enterprise framework class","l":"This neuron fires on the token just before a concrete class name is due to fill a framework slot: a ' new', the last '.' or '/' of a package path, or 'instance of the class `'. It is strongest when a setter or constructor demands a type and earlier text has named a Spring, AWS…","st":"tested","i":6,"c":6,"m":[["Apache",104],["Spring (core)",94.5],["AWS Java",91.0],["pandas",89.5],["Qt",81.5],["Rust reqwest",81.0],["no framework",81.0],["Go (halo)",79.0],["Azure .NET",75.5],["nonsense",73.0]]},{"id":"L54_N16863","k":"d","L":54,"N":16863,"t":"A neuron that fires just before February's leap day is named","l":"This neuron tracks whether the text is about adding an extra day to the calendar. Leap-year prose keeps it moderately active throughout. It peaks on the token just before February's day number, where the model is about to write the inserted day, usually 29 and sometimes Sweden's…","st":"tested","i":10,"c":4,"m":[["leap year 2024",246],["leap second",230],["leap year",204],["other 'leap'",163],["Sweden Feb 30",137],["civic control",104],["no cue",80.5],["nonsense",72.5],["common year",68.0]]},{"id":"L54_N17029","k":"d","L":54,"N":17029,"t":"The 'fung' fragment neuron: fires while the word could still be fungus","l":"This neuron fires on the spelled fragment 'fung' while the word is still unfinished, and how hard it fires tracks how likely the word is to become a fungus word. The same 'ung' token runs from about +13 after 'In banking, Mr. F' to +126 after 'Genus: F'. Complete words like…","st":"tested","i":6,"c":8,"m":[["mycology label",136],["closed word",136],["coral Fungia",103],["fungus frame",97.5],["sound-alike",49.8],["bare name",41.2],["finance name",36.5],["name fragment",28.9],["fungible copula",28.5],["plant control",0.6]]},{"id":"L54_N17443","k":"d","L":54,"N":17443,"t":"A necessity dial at the negator: owned limits up, released obligation down","l":"At a negation word, this neuron reads what the negation is about to deny. If the frame points to a limit the speaker owns or states impersonally ('Unfortunately, we are not able', 'it is not possible'), it goes positive. If the frame points to release from necessity ('X is hard.…","st":"tested","i":7,"c":7,"m":[["we, regret",25.8],["wish regret",21.0],["impersonal limit",17.4],["plain fact",9.2],["you, regret",4.7],["happy register",0.6],["does not imply",-8.2],["math need-not",-27.5],["release turn",-49.2],["how-to release",-63.5]]},{"id":"L54_N19223","k":"d","L":54,"N":19223,"t":"After 'fly' it fires; a beauty word opening a title flips it","l":"This neuron fires on the word right after 'fly' in a familiar phrase. The response lands on the 'on' in 'fly on the wall', the 'to' in 'Fly Me to the Moon' and the 'the' in 'fly in the ointment', not on 'fly' itself. Other motion verbs get a weaker version. Its negative side…","st":"tested","i":4,"c":6,"m":[["Fly capitalized",48.5],["fly (core)",36.8],["flying",18.6],["modified fly",14.1],["bee control",4.5],["swapped noun",3.9],["literal insect",3.7],["other motion verb",2.9],["nonsense",2.2],["moth control",1.1]]},{"id":"L54_N23722","k":"d","L":54,"N":23722,"t":"Sets versus structures: a signed reader of logical object kind","l":"This neuron tracks which kind of logical object the text is quantifying over and gives it a sign. Sets, Kripke frames and propositions push it positive, peaking on the word right after Russell's 'the set of all sets'. The objects of first-order model theory (structures…","st":"tested","i":6,"c":8,"m":[["set",25.1],["Kripke frame",20.6],["possible world",15.2],["proposition",13.8],["nonsense",9.4],["group (control)",2.0],["integer (control)",1.7],["formula",-20.5],["model",-21.1],["structure",-47.5]]},{"id":"L54_N24941","k":"d","L":54,"N":24941,"t":"A detector for lorem-ipsum generator words, not for Latin","l":"This neuron detects the vocabulary of lorem-ipsum generators: words like Donec, Nullam, bibendum and vulputate. Inside such text it fires on almost every token and nears its maximum of about 350. It does not respond to Latin in general, since Caesar and Vergil stay silent, and…","st":"tested","i":6,"c":4,"m":[["generator lorem",115],["Horace (shared words)",107],["English",96.5],["document start",96.5],["pseudo-words",92.5],["canonical opener",92.0],["English placeholder",88.0],["base64",5.0],["Vergil",4.4],["Caesar",-1.8]]},{"id":"L54_N5655","k":"d","L":54,"N":5655,"t":"Exposed to a viewer versus a hidden inner self","l":"This neuron runs along one signed axis. It turns positive when something normally hidden, especially a body, is put on display to an onlooker, or when one person looks at another. It turns strongly negative, about twice as deep, when text turns to a person's hidden emotional…","st":"tested","i":6,"c":7,"m":[["body bruise",22.2],["tattoo",21.5],["underwear",21.4],["object stain",14.1],["object mark",13.1],["always-seen body",7.4],["scenery",5.9],["nonsense",4.6],["plain object",4.5],["inner self",-14.0]]},{"id":"L54_N6992","k":"d","L":54,"N":6992,"t":"One gate for computer hardware; the sign says parts or kernel","l":"This neuron turns on for any computer-hardware text, then chooses a sign. Physical parts and their specs, such as memory kits, DIMM and PCIe slots, PSU connectors and CPU sockets, drive it strongly negative in any language or format. The operating system's view of a device, such…","st":"tested","i":5,"c":8,"m":[["kernel driver",12.0],["no-hardware control",0.4],["app runtime",-1.5],["car part",-2.6],["ip link eth0",-5.2],["switch as product",-21.8],["laptop spec",-51.5],["word motherboard",-55.5],["PSU connector",-93.5],["DDR4 kit",-97.0]]},{"id":"L54_N8699","k":"d","L":54,"N":8699,"t":"A neuron that sees “put them to shame” coming","l":"Fires on the rival in a playful boast: the athletes in “fleas put Olympic athletes to shame,” the GPT-4 in “leaves GPT-4 in the dust,” before the idiom is even finished. The same fact said plainly (“fleas jump higher”) stays near zero. Switch it off and “to” after “Fleas put…","st":"tested","h":"../notes/L54_N8699/","i":8,"c":8,"m":[["fleas vs. Olympians",255],["shames fancy restaurants",231],["GPT-4 in dust",192],["Chinese 相形见绌",116],["outperforms GPT-4",34.5],["put kids to bed",8.75],["plain “harder than”",0.95]]},{"id":"L55_N11647","k":"d","L":55,"N":11647,"t":"Predicts an island-group name next; negative before spaceflight and planet figures","l":"This neuron fires when the next token should name an island group. It peaks on the function word just before an archipelago name, such as 'the' before Dodecanese, and on subword pieces inside such names. Mountains, lakes and country blocs in the same frame stay near zero, and…","st":"tested","i":7,"c":7,"m":[["fictional atoll",134],["Greek island",98.5],["Ryukyu island",92.5],["large island",32.5],["coastal microstate",25.1],["mountain",4.2],["country bloc",3.6],["no place",-0.6],["lake chain",-0.8],["outer planet",-7.7]]},{"id":"L55_N11703","k":"d","L":55,"N":11703,"t":"A Colorado neuron that fires where the Colorado value comes next","l":"This neuron reads Colorado from any cue, whether the state name, a Front Range town, a university or an area code, and fires hardest on the token just before the document supplies a Colorado-specific value such as a phone number. The gate opens for any place-linked slot; the up…","st":"tested","i":9,"c":5,"m":[["Colorado",233],["Fort Collins",191],["Denver",178],["Lucerne",119],["La Paz",110],["nonsense",5.1],["Texas",-1.9],["Utah",-18.0],["Alaska",-27.9],["North Dakota",-45.5]]},{"id":"L55_N13844","k":"d","L":55,"N":13844,"t":"Onset of an expected Slavic л/тр word, negative on 'good news'","l":"This neuron fires on the first fragment of a Slavic word that begins with л or тр, and only when the context makes a Slavic word expected: the same ' л' token scores around +18 after 'In Russian' but about +3 after 'In Mongolian'. Later pieces of the word carry a weaker echo…","st":"tested","i":4,"c":6,"m":[["Russian (core)",17.2],["Romanian",15.6],["English control",13.4],["Serbian",12.6],["Ukrainian",12.4],["Bulgarian",11.8],["no-language control",3.9],["Mongolian",3.0],["Kazakh (Cyrillic, non-Slavic)",2.7],["Tajik",2.5]]},{"id":"L55_N14273","k":"d","L":55,"N":14273,"t":"Bite verbs fire it, but favored-factor products are what it computes","l":"This Qwen3-32B neuron has two unrelated faces. It fires positive on verbs of biting or clenching, such as 咬 and clenched his, but that side changes nothing downstream. Its working side is negative: at the space after '=' in a worked solution, when every factor of the coming…","st":"tested","i":10,"c":6,"m":[["same 728, pre-multiplied",2.4],["no numbers",-0.2],["half-strength pair",-1.4],["one factor vetoes",-1.6],["unfavored 729",-3.8],["addition control",-4.2],["weak favored",-8.2],["favored square",-39.5],["favored pair",-64.0],["favored chain",-178]]},{"id":"L55_N14892","k":"d","L":55,"N":14892,"t":"Ada-family declarations up, SSH and Postgres session plumbing down","l":"This neuron reads which technical dialect a document is written in and signs its output. Ada and its descendants (VHDL, Oracle PL/SQL, COBOL) push it up, whether as code or as prose about their type rules; the connection and authentication layer of client-server protocols pushes…","st":"tested","i":5,"c":6,"m":[["COBOL",40.2],["Ada (core)",34.0],["VHDL (core)",27.6],["PL/SQL",20.1],["Pascal (halo)",14.6],["Verilog (control)",1.8],["nonsense",-0.6],["Python (control)",-0.9],["MySQL (shallow neg)",-33.0],["Postgres (floor)",-46.8]]},{"id":"L55_N15120","k":"d","L":55,"N":15120,"t":"Buried-in-ash preservation, from Pompeii to Jehol, against stratigraphic catalogs","l":"This neuron fires at the moment a text is about to say how buried remains were preserved: the burial agent, depth or toll. Two scenes drive it about equally hard: Vesuvius burying Pompeii and Herculaneum, and Jehol fossils preserved in volcanic ash, both near +100. Other…","st":"tested","i":7,"c":7,"m":[["Jehol",61.5],["Pompeii",55.0],["Herculaneum",44.2],["Solnhofen",39.0],["Joya de Cerén",31.0],["Saint-Pierre",28.6],["ammonite zone",28.4],["nonsense",27.8],["Texas control",18.6],["bog body",12.5]]},{"id":"L55_N16641","k":"d","L":55,"N":16641,"t":"A neuron for 'X is really a Y in disguise'","l":"This neuron fires when a sentence says something is really a different kind of thing than it appears or is called: a setback that was an opportunity, a 'granola bar' that is a candy bar, an equation that is really a quadratic. It rises on the revealed noun and peaks on the ' in'…","st":"tested","i":8,"c":5,"m":[["opportunity (core)",155],["in-the-making",153],["blessing (idiom)",134],["gift",116],["same valence",68.0],["taxonomy",61.5],["no contrast",57.2],["literal costume",54.8],["nonsense",27.2]]},{"id":"L55_N16859","k":"d","L":55,"N":16859,"t":"A hangover neuron that predicts 'hangover' and opposes fibromyalgia","l":"This neuron tracks one kind of misery: the morning after drinking. It fires hardest on the word just before 'hangover' and directly promotes writing ' hang'. Knocking it out cuts that prediction from about two-thirds to under a tenth. Headaches, migraines, and figurative…","st":"tested","i":8,"c":7,"m":[["office party",178],["heavy drinking",140],["figurative",114],["word 'hangover'",104],["migraine",87.0],["back pain",83.0],["jet lag",65.5],["flu",57.0],["no-malaise control",45.0],["fibromyalgia",-58.8]]},{"id":"L55_N1747","k":"d","L":55,"N":1747,"t":"A Jurassic Park neuron: dinosaur cloning premise, against other prehistory","l":"This neuron tracks whether the prehistoric subject is dinosaurs, and it is strongest when the Jurassic Park premise is in play. The film frame and the dinosaur subject add together: either one alone gives a partial response, and both push slot tokens like 'from' and 'trapped in'…","st":"tested","i":8,"c":7,"m":[["dinosaurs",122],["T. rex",104],["raptors",92.0],["pterosaurs",83.0],["trilobites",78.5],["dodos",71.5],["no-animal control",71.5],["hominids",39.5],["mammoths",36.0],["antiquity",31.6]]},{"id":"L55_N18366","k":"d","L":55,"N":18366,"t":"A chili-heat neuron that fires where the pungency comes next","l":"This neuron tracks chili heat as the live topic. Throughout chili or capsaicin text it stays moderately on, and it spikes where a heat level, a Scoville value or the spicy sense of 'hot' should come next. It does this even without food words: a list of senses of 'hot' lights it…","st":"tested","i":8,"c":6,"m":[["capsaicin",145],["habanero",110],["nonsense",53.8],["bell pepper",42.8],["menthol",42.5],["ginger",36.8],["wasabi",35.8],["tomato",31.5],["black pepper",30.1]]},{"id":"L55_N20495","k":"d","L":55,"N":20495,"t":"Fires just before a character's disguise is named, strongest across gender","l":"This neuron fires at the point where a text is about to say that someone is appearing as someone else, such as 'the lawyer was actually Portia in…' just before 'disguise'. Literary reveals, scammers, undercover posing and costumes all trigger it. Disguises that cross gender push…","st":"tested","i":8,"c":8,"m":[["woman as man",166],["king disguised",137],["queen as man",133],["man as man",105],["undercover",74.5],["costume",67.0],["impostor",57.5],["no disguise",55.0],["plain control",16.9],["nonsense",4.2]]},{"id":"L55_N21606","k":"d","L":55,"N":21606,"t":"A relay-race neuron that predicts the baton before it appears","l":"This neuron tracks relay races. Once a text is about a relay (a 4x100, a torch relay, a baton passed leg to leg), it fires on the word just before the baton or the relay event should appear, and it pushes the model to predict that word. Shift and power handovers give weaker…","st":"tested","i":8,"c":7,"m":[["4x100 relay",201],["'relay' alone",159],["nurse shift",155],["family firm",118],["torch relay",112],["dinner",81.5],["royal reign",68.5],["nonsense",54.0],["volleyball",42.8]]},{"id":"L55_N2214","k":"d","L":55,"N":2214,"t":"Getting ready to say “did it anyway”","l":"Fires on the last word of something done despite a reason not to, one word before “anyway”: “they knew it was wrong, but they did it”. Zeroing it drops “anyway” after “…but the kids jumped in” from 95% to 7%, though on other sentences the drop is often smaller. Same slot in…","st":"tested","h":"../notes/L55_N2214/","i":9,"c":9,"m":[["BUT DO IT",274],["Italian",260],["sour grapes",208],["drove down it",174],["heading that way",100],["obeyed warning",13.2],["at any rate",-66.5]]},{"id":"L55_N23033","k":"d","L":55,"N":23033,"t":"Land, seen from the water","l":"Fires just before a text says 'land' in the sailor's sense: Noah's raven sent 'to find dry land', the Pinta crewman who 'spotted land', hurricanes that 'hit land'. Land you buy, land a plane or land a job leave it silent. Its negative side dips when a text heads the other way…","st":"tested","h":"../notes/L55_N23033/","i":8,"c":8,"m":[["Noah's raven",206],["Pinta lookout",175],["German sailors",144],["over the ocean",125],["in the jungle",9.88],["acres of land",0.32],["interior towns",-85.5]]},{"id":"L55_N24982","k":"d","L":55,"N":24982,"t":"A slot neuron for where sea creatures live on the shore","l":"This neuron marks the moment a text is about to say where a sea creature lives vertically. It fires on the small word just before the shore level or depth is named, as in 'found from the [lower shore]', and it helps predict words like 'intertidal' or 'lower'. Shore context alone…","st":"tested","i":6,"c":8,"m":[["limpet, rocky shore",203],["chiton, coast",165],["crab, sand beach",149],["anemone, ocean",84.0],["wreck, no organism",55.5],["garden rose",46.8],["river trout",30.8],["alpine plant",22.5],["lake perch",18.8],["no habitat",13.3]]},{"id":"L55_N25011","k":"d","L":55,"N":25011,"t":"Anticipating 'past life': a slot neuron for prior lifetimes","l":"This neuron predicts an upcoming 'past life', 'previous life' or 'former lives' phrase when the text frames a person as having lived before. Examples are a psychic claim, a Jataka tale about the Buddha, or the idiom 'in a past career'. It builds over the lead-in and peaks on 'in…","st":"tested","i":8,"c":6,"m":[["psychic claim",328],["belief claim",237],["Jataka Buddha",157],["soulmate",97.5],["reborn-as",89.0],["former career",42.8],["stage role",31.6],["nonsense",27.1],["object incarnation",8.5],["no frame",-1.7]]},{"id":"L55_N3087","k":"d","L":55,"N":3087,"t":"Explained mimicry, peaking just before the imitator is named","l":"L55/N3087 fires when a text explains one thing copying or passing for another: harmless snakes mimicking coral snakes, syphilis 'the great imitator', people impersonating police. It peaks on the token just before the imitator or the imitation word is named, and it pushes words…","st":"tested","i":7,"c":6,"m":[["pretends",80.0],["Batesian mimic",72.5],["copies",71.0],["mimics",62.5],["control",37.5],["looks like",32.2],["shares woods",29.6],["mistaken for",27.1],["no relation",24.0]]},{"id":"L55_N4037","k":"d","L":55,"N":4037,"t":"One garden axis: summer moth flowers positive, spring clover meadow negative","l":"This neuron rates a passage on one garden axis. Positive is the summer moth garden: Datura and other night-opening trumpet flowers, hawk moths and noctuids, pumpkins and squash. Negative is the spring meadow and orchard: clover and pasture grass, bees and aphids, small blue…","st":"tested","i":5,"c":4,"m":[["Cucurbita",91.5],["Datura core",85.0],["macro-moth",85.0],["night bloomer",79.5],["summer annual",60.8],["day Solanaceae",41.8],["bee (neg)",-7.8],["no taxon",-8.8],["pasture (neg)",-22.8],["blue (neg)",-35.5]]},{"id":"L55_N619","k":"d","L":55,"N":619,"t":"A zodiac date-span slot neuron that runs negative on feng shui","l":"This neuron marks the moment a horoscope is about to state a sign's dates. It fires hardest on the slot where the day number follows the month, as in 'Taurus: April ▯20'. Any one cue is enough to set this up: a sign name, a 'born between' frame, or the familiar 20/21 cutoff…","st":"tested","i":6,"c":6,"m":[["sign season",70.5],["sign label",50.0],["single sign date",37.8],["astronomy",34.0],["birth frame",32.2],["zodiac cutoffs",24.4],["mundane control",1.1],["shipping control",-0.2],["arbitrary cutoffs",-0.9]]},{"id":"L55_N8722","k":"d","L":55,"N":8722,"t":"A Thor detector whose volume is turned up by money talk","l":"This neuron multiplies two signals. Letter units detect a word-initial Thor token and set the sign: positive on Thor and Th-names, negative where a market price or dataframe value is about to be read. Money or economics context sets the gain and deepens both poles. 'The…","st":"tested","i":4,"c":5,"m":[["inflation+economist Thor",72.5],["economist Thor",63.5],["statistician Thor",53.0],["programmer Thor",36.2],["physicist Thor",31.4],["economist Thurston",28.0],["poet Thor",25.9],["poet Thurston",6.0],["economist Thompson",2.7],["economist Lars",2.7]]},{"id":"L56_N1041","k":"d","L":56,"N":1041,"t":"Specific body part ahead versus the whole system","l":"This neuron is a signed slot detector in anatomy and nature writing, read on the word just before a noun. It goes positive when the next word should be a specific named part, such as a bone, vessel or shore zone. It goes negative when the next word should be the containing…","st":"tested","i":5,"c":5,"m":[["lymph network",37.8],["licence list",25.1],["bone shaft",16.6],["shore zone",15.7],["nonsense",1.3],["motor control",-15.4],["street control",-15.4],["belong to",-31.5],["organs of",-43.0],["sistema",-96.0]]},{"id":"L56_N11540","k":"d","L":56,"N":11540,"t":"The slot before bare 'town': writes the 'in town' idiom","l":"This neuron fires on the word just before a bare 'town': the ' in' of 'the only game in', the ' to' of 'comes to', the ' of' of 'the rough part of'. It writes 'town', and ablating it collapses P(town) from near-certain to near zero. It is not a general P(town) detector. Slots…","st":"tested","i":8,"c":8,"m":[["new X in",213],["part of",156],["only game in",147],["comes to",110],["newcomer block",78.0],["superlative",63.5],["arrival, city",39.8],["P(town) high",4.7],["numeric bound",4.5],["other bare noun",-1.4]]},{"id":"L56_N1157","k":"d","L":56,"N":1157,"t":"Leading 4 up, leading 6 down; Korea opens the gate","l":"L56/N1157 watches the first digit of a number the model is writing. A leading 4 in a figure it just computed pushes the neuron up; a leading 6 in a year, a time constant, a phone number or a decimal pushes it down. Plain counts stay silent. Korea acts as a gate: the word Korean…","st":"tested","i":4,"c":6,"m":[["Manchuria",66.5],["Korean province",66.0],["Silla Korea",65.5],["Texas",65.0],["Korean",64.5],["Joseon",63.0],["nonsense",62.8],["Japan",62.0],["Mongolia",61.2],["Bolivia",61.0]]},{"id":"L56_N11866","k":"d","L":56,"N":11866,"t":"Predicts the bubbly-foam word (lather, suds, froth, head) before it arrives","l":"This neuron predicts foam. It fires on the token just before a word like lather, suds, foam, froth, a beer head or espuma is due, in any language. It peaks in hard-water and shampoo text, where lather is most predictable. Ablating it cuts the model's foam prediction sharply…","st":"tested","i":8,"c":8,"m":[["beer",208],["egg whites",186],["dish soap",157],["shampoo",137],["latte milk",120],["sea foam",95.0],["clotted cream",60.0],["nonsense",57.5],["oil control",50.2],["paint (texture)",43.2]]},{"id":"L56_N12242","k":"d","L":56,"N":12242,"t":"A slot neuron for names of species lost to extinction","l":"This neuron predicts that the name of an extinct species is coming next. It fires on the word just before that name, in text about bringing species back or about humans wiping them out. The frame opens the slot even before any animal is named: 'scientists might be able to bring…","st":"tested","i":7,"c":7,"m":[["de-extinction",216],["overkill",195],["admixture",142],["asteroid",71.5],["fossil lineage",52.8],["endangered",30.5],["dead language",17.4],["company control",0.7],["media control",0.7],["hominin lineage",-78.5]]},{"id":"L56_N13762","k":"d","L":56,"N":13762,"t":"Negative pole reads fate off a chart: astrology and credit predict the noun","l":"This unit works through its negative pole. It fires strongly negative whenever a person's fate or standing is being read off an assessment system: natal charts and horoscopes in any culture, credit reports and FICO scores, and more weakly tarot, numerology and percentiles.…","st":"tested","i":9,"c":6,"m":[["bank (positive)",14.9],["medical control",6.2],["nonsense",0.3],["game control",-3.0],["screening",-72.0],["tarot halo",-75.5],["numerology halo",-85.5],["credit core",-127],["astrology core",-170],["Vedic",-177]]},{"id":"L56_N16113","k":"d","L":56,"N":16113,"t":"Surface features versus sieve grades: a signed neuron for small sizes","l":"This neuron fires when a text is about to give the size of small things, and its sign says what kind. Positive means features built into a surface, such as moth-eye pillars, shark-skin riblets, gecko hairs or pores in a film, and it peaks just before the number. Negative, and…","st":"tested","i":7,"c":6,"m":[["moth-eye pillars",90.0],["nanocones (core)",69.0],["film pores",63.0],["leaf hairs (halo)",32.8],["no-cue control",10.9],["bare nanoparticles",9.4],["macro control",8.7],["coffee grains",8.6],["sifted flour",-18.4],["sieved powder",-53.8]]},{"id":"L56_N16469","k":"d","L":56,"N":16469,"t":"A vitamin C neuron that writes 'scurvy', with folate as its opposite","l":"This neuron tracks whether the text is about vitamin C. Plain vitamin C prose keeps it positive. It peaks right before 'scurvy' or 'ascorbic' is written, and ablating it there wipes out that next word. Deficiency diseases of other vitamins get weaker spillover, and sailor or…","st":"tested","i":7,"c":7,"m":[["vitamin C",282],["citrus",276],["ascorbate",272],["vitamin D",196],["no nutrient",179],["thiamine",157],["B12",135],["folate",128],["niacin",120],["iodine",116]]},{"id":"L56_N17041","k":"d","L":56,"N":17041,"t":"Predicting a body about to roll into a ball","l":"This neuron predicts that a living body is about to roll itself up, and it pushes 'roll' as the next word. It fires hardest just before the verb for pill bugs, armadillos, pangolins and hedgehogs. It fires about half as much for other animal defenses, people dropping and…","st":"tested","i":8,"c":6,"m":[["pill bug",112],["hedgehog",88.0],["armadillo",84.0],["ladybug",71.0],["opossum (other defense)",58.5],["dog",33.5],["skunk (other defense)",29.2],["crane fly",19.8],["no-body control",11.6],["long antennae",-22.1]]},{"id":"L56_N17420","k":"d","L":56,"N":17420,"t":"A signed slot: a birdwatcher's paradise versus the literal hell","l":"L56/N17420 fires just before a word like paradise, heaven or hell, and its sign says which sense is coming. It goes strongly positive when a worldly place or thing is about to be called someone's realm, peaking on the 's of 'a birdwatcher's' (about +150). It goes negative when a…","st":"tested","i":9,"c":7,"m":[["birdwatcher's",117],["pedestrian's (hell)",115],["spammer's",109],["ordinary user",82.0],["idiom",70.5],["breeding ground",45.5],["animal audience",32.8],["secular control",4.1],["literal heaven",-9.9],["literal hell",-26.2]]},{"id":"L56_N17789","k":"d","L":56,"N":17789,"t":"A Zeno's-paradox neuron that writes 'Achilles', and is negative for Descartes","l":"This neuron tracks whether the model has recognized Zeno's paradoxes of motion. Zeno's name and halving-distance content each raise it on their own and add up when combined. At linking words such as 'paradox of' it pushes the model toward 'Achilles' and 'Dichotomy', and ablating…","st":"tested","i":8,"c":6,"m":[["Zeno (core)",192],["bare Zeno",188],["Stoic Zeno",184],["Eleatic",115],["classical",56.5],["no-cue control",25.5],["other paradox",25.4],["early modern",24.9],["Descartes",-23.1]]},{"id":"L56_N18200","k":"d","L":56,"N":18200,"t":"Italian essere before a participle saying something happened to the subject","l":"This neuron fires on an Italian form of essere at the moment the model expects a past participle saying something happened to the subject: the ship sank, the session was closed. It reads the event, not the syntax. Reflexive perfects and property copulas stay flat, agentive…","st":"tested","i":6,"c":8,"m":[["sarebbe",61.2],["dative mi è",51.0],["è (core)",48.0],["era",43.8],["English was",9.4],["Spanish ha",1.3],["no auxiliary",1.1],["reflexive si è",0.9],["German ist",-3.4],["fu",-5.5]]},{"id":"L56_N20753","k":"d","L":56,"N":20753,"t":"The neuron that fires on the space+0xD0 byte before й","l":"Qwen's tokenizer keeps most space-prefixed Cyrillic letters whole, but splits й, Й, Щ and extended capitals such as І, Є, Ј into a space-plus-lead-byte token and a second byte. This neuron fires almost only on that first token, near 200 in Slavic prose, before the letter is…","st":"tested","i":4,"c":8,"m":[["й (0xD0 byte)",161],["Й (0xD0 byte)",161],["Щ (0xD0 byte)",161],["Ukrainian І",161],["щ (0xD1 byte)",43.0],["ю (0xD1 byte)",43.0],["English web",12.6],["whole-token и",1.9],["Latin Y",-0.1],["whole Cyrillic word",-0.5]]},{"id":"L56_N20810","k":"d","L":56,"N":20810,"t":"Predicting an 'on' that lands on someone","l":"This neuron fires on the word just before an 'on' or 'upon' that is about to land on a target, and it is strongest when the target is a person: an irony 'not lost on' someone, a crowd that 'turned on' its leader, a petition that 'calls on' States. It is silent for topic 'on'…","st":"tested","i":8,"c":8,"m":[["turned (attack)",74.5],["descended",58.2],["imposed fine",23.9],["called (appeal)",16.4],["effect, projected",4.3],["relied (halo)",3.6],["no-on control",-1.1],["focused (topic)",-1.5],["measured effect",-7.0],["felt pressure",-8.9]]},{"id":"L56_N21500","k":"d","L":56,"N":21500,"t":"The verb that opens small talk: 'tried to make [conversation]'","l":"This neuron fires on the verb that opens the phatic-talk idiom: 'make' in 'tried to make small talk', or 'start' before 'a conversation'. It fires when the scene calls for talk that only fills a social gap, such as an awkward silence, a stranger, a crush or an etiquette lesson.…","st":"tested","i":7,"c":8,"m":[["stranger, silence",201],["dinner party",171],["relaxed chat",147],["crush",129],["old friend",51.0],["object, no talk",45.5],["nonsense",30.2],["kids at play",25.1],["dance history",-43.5]]},{"id":"L56_N2175","k":"d","L":56,"N":2175,"t":"Version-slot detector in software names, signed by Unix lineage","l":"This neuron marks the spot inside a software identifier where a version or tag is about to follow: the colon in alpine:latest, the paren in fstab(5), the digit in nginx:1.25. Remove the continuation and it goes quiet. The name decides the sign. Alpine, BSD, Debian, nginx and…","st":"tested","i":6,"c":4,"m":[["FreeBSD",28.9],["Alpine (core)",28.6],["PostgreSQL",25.1],["nginx",11.7],["Ubuntu (halo)",5.1],["no-name control",2.9],["Node (near zero)",-2.1],["Java (near zero)",-4.7],["CentOS (negative)",-14.2],["Oracle Linux (negative)",-35.8]]},{"id":"L56_N21766","k":"d","L":56,"N":21766,"t":"A neuron that predicts 'long' after 'won't last for'","l":"This neuron fires when a sentence says something won't last, as in 'that hive won't be empty for'. It peaks on the word just before 'long' would appear. Negation is required: 'the calm will last' leaves it silent. When active, it pushes the model toward extent words like 'long'…","st":"tested","i":7,"c":6,"m":[["core, stay",298],["core",164],["past tense",162],["no 'for'",134],["weak slot",81.0],["rule-like",28.2],["affirmative",7.2],["bare negation",2.3],["no cue",-0.3],["affirmative duration",-2.7]]},{"id":"L56_N2209","k":"d","L":56,"N":2209,"t":"A \\prime detector that fires hardest when the prime is unannounced","l":"This neuron in Qwen3-32B is a signed detector of prime marks. On the positive side it fires on the LaTeX token 'prime' after a backslash in math, and its size tracks whether the mark is news: a primed symbol dropped in without warning fires about twice as hard as one arriving…","st":"tested","i":6,"c":8,"m":[["finance control",87.0],["lone symbol",81.5],["nonsense control",80.5],["integral control",72.5],["primed frame",57.0],["arc minutes",50.8],["derivative",48.8],["Chinese derivative",48.5],["tangent slope",45.0],["differentiate",44.8]]},{"id":"L56_N23633","k":"d","L":56,"N":23633,"t":"Signed category slot: process or stage versus organism interaction","l":"This neuron fires on the token just before a text names what kind of thing something is, such as ' as' after 'Classify the process' or the separator in a type table. Its sign depends on the kind being named. Transformations, modes and stages push it positive. Interactions…","st":"tested","i":5,"c":6,"m":[["adiabatic gas",35.5],["grief stages",34.2],["heat flow",31.4],["Na + Cl2",24.9],["nonsense control",22.9],["metamorphosis (halo)",21.2],["electrons repel",15.2],["static shape",5.5],["predation",-0.8],["symbiosis",-37.0]]},{"id":"L56_N2458","k":"d","L":56,"N":2458,"t":"The number that 'Of the' introduces, and its two negative mirrors","l":"This neuron fires on the number that states the whole of a partitive, but only when English 'Of' or 'Out of' opens the clause: 'Of the 200' lights up on the digits, while 'In the 200' stays dark. It also fires inside remembered I/O signatures like sprintf. Its negative side is…","st":"tested","i":4,"c":3,"m":[["Out of the",20.5],["Of the (core)",16.2],["German Von",7.2],["back-reference",6.4],["Percent of (halo)",5.7],["plain quantity",5.3],["Half of (halo)",4.2],["For (swapped prep)",0.4],["nonsense",-0.3],["In (swapped prep)",-0.6]]},{"id":"L56_N25397","k":"d","L":56,"N":25397,"t":"A signed dial for 'for': promotes it after ground words, suppresses after transactions","l":"L56/N25397 is a signed dial on the next-token decision about 'for', across languages. Its strong negative pole fires on words that attribute a ground, credit or reason and must take 'for' next: responsible, known, grateful, 'atomos is Greek'. Zeroing it lowers the probability of…","st":"tested","i":6,"c":6,"m":[["pay 50 cents",8.0],["treated (halo)",6.3],["substitute (word)",5.3],["searching (purpose)",3.9],["arrested",0.5],["no-for control",-0.2],["grateful",-3.9],["best known",-27.0],["responsible (core)",-46.2],["Greek gloss",-63.2]]},{"id":"L56_N4987","k":"d","L":56,"N":4987,"t":"Att- stems of the classical world, and the composers that suppress them","l":"This neuron fires on an unfinished Att- stem, the start of Attic, Attica, Atticus or Attalus, in any case and even in Greek script. Classical context multiplies the response: the stem alone is moderate, but in vase catalogues, Latin name lists or Attic grammar it climbs past…","st":"tested","i":4,"c":9,"m":[["Latin names",85.0],["vase catalogue",76.0],["Attalids",75.5],["Cicero",67.0],["neutral",62.5],["Greek history",54.0],["Debussy",54.0],["modern home",47.2],["Attlee era",29.9],["Penderecki",-77.5]]},{"id":"L56_N6203","k":"d","L":56,"N":6203,"t":"An unfinished-word neuron: a-stems push up, com-p stems push down","l":"This neuron fires on a token that leaves a word unfinished in a slot expecting a content word, and reports which family the letters so far belong to. Stems beginning with 'a' plus a stop (mis-ap|plied, un-ap|ologetic, ak|nowledge) drive it strongly positive; stems reading com-p…","st":"tested","i":4,"c":6,"m":[["prefixed a-stem",62.0],["a+k stem",58.8],["a-stem core",41.5],["s-onset",32.2],["other vowel",22.0],["no-fragment control",-0.0],["whole word control",-0.4],["Tagalog kump",-14.6],["consonant e-stem",-14.9],["com-p",-18.6]]},{"id":"L56_N6360","k":"d","L":56,"N":6360,"t":"Predicts the 'we're on it' reply once a problem reaches its handler","l":"This neuron predicts the 'we're on it' predicate. It fires on the word that launches what the responsible party is doing about a matter raised with them: 'the developers [are]', 'we've got our best [people]', 'we're [on] it'. Two conditions set its strength: the problem must…","st":"tested","i":8,"c":8,"m":[["best people",93.0],["developers",77.0],["we",36.8],["sadly, devs",36.2],["MIT researchers",33.8],["nonsense",10.2],["parents",9.9],["customers",7.5],["pages",2.7]]},{"id":"L56_N6442","k":"d","L":56,"N":6442,"t":"A switch-list neuron: flag letters after a declared Options header","l":"This neuron marks the name slot of an entry in a list that the context has declared to be command-line switches, for example the 'a' in '-a, --all' under a Usage or Options header. The frame matters far more than the entry: '-a toothbrush' fires near the top under 'Usage: pack…","st":"tested","i":8,"c":7,"m":[["baseline table",29.9],["Flags header",14.6],["Usage+Options",13.4],["inline stats",9.9],["man OPTIONS",9.1],["env-var list",5.1],["prose flag",4.6],["shortcuts",1.9],["packing list",0.6],["no cue",0.2]]},{"id":"L57_N14403","k":"d","L":57,"N":14403,"t":"A neuron that predicts the word for laziness before it arrives","l":"L57/N14403 fires at the moment the next word should mean lazy, sloth or listless, in any language. It fires before the word appears, not on it. Synonym lists, translations and confessions like 'not tired, just…' drive it to about +400. Looser slack-state contexts give weaker…","st":"tested","i":6,"c":8,"m":[["lazy (core)",270],["slothful",262],["lethargic",236],["sluggish (halo)",217],["idle",107],["tired (fatigue)",87.5],["sleepy",68.0],["antonym",10.0],["nonsense",8.2],["control",0.8]]},{"id":"L57_N16430","k":"d","L":57,"N":16430,"t":"Expects a hand-configuration word; strongest in sign-language entries","l":"This neuron fires where the text is about to specify how a human hand is configured, such as its shape, palm orientation or movement. It peaks around +200 in sign-language dictionary entries and parameter lists, and its output pushes toward words like 'palm' and 'fingers'.…","st":"tested","i":7,"c":6,"m":[["ASL sign",163],["other sign language",161],["sign topic, no hand",144],["clinical hand",102],["foot/gait",100],["dance mudra",79.0],["nonsense",72.5],["robot hand",69.5],["non-body object",68.5],["gripper",63.2]]},{"id":"L57_N16863","k":"d","L":57,"N":16863,"t":"A 'preview word coming' neuron across languages","l":"This neuron signals that a word for a brief look is about to be written, such as glimpse, sneak peek, preview, first look, aperçu or 预览. It peaks on the slot just before that word and on its unfinished pieces, and it drops once the word is complete. The same 'pre' fragment turns…","st":"tested","i":6,"c":8,"m":[["live UI",52.0],["teaser",41.0],["trailer",23.4],["sample",10.9],["control",9.7],["forecast",4.8],["nonsense",-1.2],["precursor",-5.6],["ancestor",-8.4],["precancerous",-18.2]]},{"id":"L57_N19496","k":"d","L":57,"N":19496,"t":"Domino neuron: tiles and cascades, negative on chess ratings","l":"This neuron responds to dominoes as tiles. It fires on 'fell like dominoes', on 1x2 dominoes tiling a chessboard, and on the table game, and it extends more weakly to checkers and checkerboard reasoning. Its causal effect lands on the slot before the word, where it helps predict…","st":"tested","i":7,"c":6,"m":[["banks (cascade)",106],["regimes (core)",98.0],["ratings",92.5],["tiles (tiling)",86.5],["chess topic",76.0],["blocks",60.0],["checkers (halo)",50.0],["nonsense",49.5],["leaves (control)",22.9]]},{"id":"L57_N20004","k":"d","L":57,"N":20004,"t":"Under siege, the moment before it falls or holds","l":"This neuron tracks a thing under sustained attack, such as a fortress under siege, a crop hit by disease or a problem that resists solution. It fires on the word just before that thing's fate is stated, whether it will fall or hold. It writes 'fall/succumb'. Zeroing it at a…","st":"tested","i":8,"c":8,"m":[["besieged fortress",79.5],["encircled city",65.0],["crop vs disease",37.0],["garrison commander",31.1],["attacking army",15.9],["tower vs army",12.3],["tower vs storm",7.2],["no adversary",4.6],["no-contest control",-0.4]]},{"id":"L57_N2065","k":"d","L":57,"N":2065,"t":"An arithmetic-subject neuron that fires at connectives before technical terms","l":"This neuron tracks whether a passage is about arithmetic: primes, number fields, roots of unity, Galois groups, elliptic curves, Langlands. The signal builds over the context and is read out at small connectives just before the next technical term, so rich paragraphs score far…","st":"tested","i":5,"c":5,"m":[["arithmetic core",33.5],["NT name only",27.5],["'number theory'",26.1],["Langlands",21.9],["math halo",19.4],["neighbour algebra",14.6],["probability",10.6],["nonsense",3.0],["non-math control",-0.6],["Indian institution",-5.3]]},{"id":"L57_N21057","k":"d","L":57,"N":21057,"t":"Wild big-game country reads positive, catalogued countryside reads negative","l":"This neuron fires on the token before a text names or measures a piece of land, and its sign says what kind of land is coming. Grizzly, wolf and wolverine country in the Northern Rockies, Cascades and Alaska drives it strongly positive; tropical megafauna corridors and salmon…","st":"tested","i":7,"c":8,"m":[["Northern Rockies",72.5],["North Cascades",62.8],["Alaska",45.0],["Amazon",16.8],["Swiss Alps",16.8],["Sussex downs",16.8],["no place",16.8],["Michigan",-20.0],["central Texas",-20.2],["Adirondacks",-22.4]]},{"id":"L57_N22325","k":"d","L":57,"N":22325,"t":"Predicts bursting and outbreak verbs; negative on technical explosives vocabulary","l":"This neuron fires on the token just before a text says something bursts, snaps or breaks out: a person whose anger has been swallowed, a crowd before a revolt, a balloon or seed pod swollen to its limit. Its value tracks how strongly a burst verb is expected next, peaking near…","st":"tested","i":7,"c":6,"m":[["seed pod",64.5],["revolt",51.8],["emotion",40.8],["balloon",40.5],["boiler",38.0],["neutral control",38.0],["volcano (halo)",35.5],["TNT",19.0],["bridge",-7.9],["Big Bang",-19.9]]},{"id":"L57_N22809","k":"d","L":57,"N":22809,"t":"A signed slot: youth kept versus things left to mature","l":"This neuron predicts what time will do next. It fires strongly on the word just before 'young' or 'youth' in anti-aging talk, such as 'this serum keeps [you]' or 'fountain [of]'. It fires negatively when the next word should describe time finishing something, such as grout…","st":"tested","i":7,"c":8,"m":[["fountain of",157],["serum keeps you",73.5],["literal age",33.0],["keeps skin",29.8],["keeps car",15.0],["library fountain",5.7],["control",-0.4],["cheese matures",-60.8],["grout cures",-61.0],["pears ripen",-62.8]]},{"id":"L57_N2461","k":"d","L":57,"N":2461,"t":"A continue-slot neuron: resuming stories fire positive, persisting states negative","l":"This neuron fires on the token just before a word meaning 'continue' is due. What is continuing sets its sign. When a text or activity goes on or resumes, it fires positive; the 'To Be Continued' card after a cliffhanger peaks near +245. When summarizing or forecast prose says a…","st":"tested","i":7,"c":8,"m":[["interrupted",30.2],["plain resume",13.9],["sequence",11.0],["serial frame",4.9],["keep-reading",2.8],["nonsense",0.8],["legacy",-4.5],["forecast",-6.7],["summary persist",-17.5]]},{"id":"L57_N2471","k":"d","L":57,"N":2471,"t":"A slot neuron that predicts 'you will find'","l":"This neuron sits on the slot just before a clause continues and signals what kind of continuation is coming. It fires strongly negative on a subject about to be told what it will find or receive, as in 'Below [you] will find'. That negative firing raises the probability of '…","st":"tested","i":6,"c":7,"m":[["narrative control",-5.5],["nonsense",-5.7],["if-steps",-14.0],["find below",-14.0],["once-clause",-17.8],["ask any X",-37.8],["following pages",-41.5],["this folder",-49.0],["attached PDF",-51.0],["Below",-59.0]]},{"id":"L57_N24736","k":"d","L":57,"N":24736,"t":"A neuron that waits for the lullaby","l":"This neuron tracks one situation: someone is about to sing to soothe a listener in their care. Lullabies to babies are the clearest case, but a puppy or an elderly mother works too. The state builds over caregiving words and peaks where the song is named or begins. Caregiving…","st":"tested","i":6,"c":6,"m":[["puppy",145],["newborn",141],["toddler",141],["coworker",132],["elderly mother",128],["unborn baby",116],["reading-level cue",115],["sailors",102],["object",87.0]]},{"id":"L57_N2699","k":"d","L":57,"N":2699,"t":"A slot neuron predicting 'disguised' or 'concealing' comes next","l":"This neuron fires on the token just before a word of disguise or concealment is expected, such as 'a thinly[ veiled]', 'a facade[ concealing]' or 'sweaters[ to] hide'. How strongly it fires tracks how confident the model is that such a word comes next, not the topic: political…","st":"tested","i":6,"c":8,"m":[["euphemism",153],["sham politics",144],["emotional mask",137],["facade",120],["satire target",112],["literal cover",111],["nonsense",97.0],["sliced control",92.5],["thin layer",85.0],["sliced control",54.8]]},{"id":"L57_N4212","k":"d","L":57,"N":4212,"t":"A neuron that helps spell 'foreshadow' once omens appear","l":"This neuron anticipates a word of prefiguration. After a text lays out an omen, a plot setup or an early precedent, it fires on the first fragment of 'foreshadow', 'presage', 'bode' or 'harbinger' while the word is still incomplete, and it helps the model finish that word. It…","st":"tested","i":8,"c":8,"m":[["forecast signal",163],["literary setup",148],["weather omen",132],["legacy",120],["present diagnosis",110],["bird omen",108],["past sign",92.5],["numeric",83.5],["neutral noun",74.5],["no cue",59.0]]},{"id":"L57_N5889","k":"d","L":57,"N":5889,"t":"A 'CAPTCHA comes next' neuron with a 'talk with me' negative pole","l":"This neuron fires on the token just before text is expected to name a challenge that stands between a user and access. CAPTCHAs and their vendors (reCAPTCHA, hCaptcha) are the main case: peaks near +280 land on list commas and on 'without [solving]', not on the word CAPTCHA. Its…","st":"tested","i":8,"c":7,"m":[["bot-defence vendor",212],["CAPTCHA word",178],["generic gate",145],["checkout",139],["math puzzle",123],["2FA halo",118],["nonsense",108],["benign web",105],["sign-up page",100],["non-web control",94.0]]},{"id":"L57_N6278","k":"d","L":57,"N":6278,"t":"The neuron that knows a game is about to roll dice","l":"This neuron marks the moment in a game where the next move is to call for a roll or name the die: after a check-worthy action, at 'you need to', or at 'roll a'. It needs a game frame. The same action in novel prose leaves it silent. At these points it pushes 'roll' and 'd'…","st":"tested","i":8,"c":7,"m":[["DM prompt",140],["dice math",133],["nonsense",133],["TTRPG",130],["wheel randomizer",125],["no frame",116],["novel prose",96.0],["bowling",74.5],["board game",71.5]]},{"id":"L57_N6932","k":"d","L":57,"N":6932,"t":"A neuron for the open slot in 'turn the tables'","l":"This neuron fires at the open slot of the reversal idiom: the ' the' after 'turn', where 'tables' or 'tide' should come next. It fires most strongly when the story has an underdog, captive or prey about to gain the upper hand. It responds to the idiom slot even when 'tide' is…","st":"tested","i":7,"c":7,"m":[["finally turn",126],["turn the",105],["flip the",40.0],["reverse the",40.0],["swing the",26.5],["tell the",5.2],["neutral verb",4.8],["on the",-2.3],["hold back",-3.1],["tack to",-12.4]]},{"id":"L57_N7653","k":"d","L":57,"N":7653,"t":"Receiving care versus delivering it: a two-sided mental-health neuron","l":"This neuron reads mental-health and disability care and tells you which side of it a text is on. It goes positive for the people receiving care and what they are treated with, and peaks just before the next antipsychotic or study arm is named. It goes equally negative for the…","st":"tested","i":7,"c":6,"m":[["special needs",50.0],["stroke (halo)",39.0],["care settings",38.0],["psych patients",31.1],["schizophrenia",27.5],["law clerks",22.1],["chem students",7.8],["cakes (control)",-7.2],["psych interns",-16.4],["school psych",-54.5]]},{"id":"L57_N800","k":"d","L":57,"N":800,"t":"A slot for the verdict on a stance, and it writes 'mixed'","l":"This neuron fires on the word just before a verdict on a stance is spoken: how the public reacted, what critics thought, whether the research or the news is good. It works in every language tested and needs the verdict still open; one-sided prior evidence halves it and the…","st":"tested","i":9,"c":5,"m":[["my feeling",54.0],["the news",54.0],["critics' verdict",48.8],["public reaction",43.5],["sentiment",38.0],["the research",15.6],["election result",13.0],["category label",8.4],["cost (control)",7.2],["date (control)",7.2]]},{"id":"L57_N8980","k":"d","L":57,"N":8980,"t":"One geoscience gate, two signs: tectonic process versus source and shore","l":"In earth-science prose this neuron fires at the function word that opens a noun slot. It goes positive when the frame is tectonic: a rupture, plate motion, a trench or basin, peaking at 'the direction of' inside an earthquake report at around +100. It goes negative when the same…","st":"tested","i":5,"c":4,"m":[["fault zone",48.2],["Nazca plate",27.1],["basin at rest",26.1],["plateau halo",25.2],["large peninsula",9.5],["non-earth control",2.5],["nonsense place",-11.9],["regional coast",-29.0],["small headland",-32.8],["Irish headland",-65.0]]},{"id":"L58_N11554","k":"d","L":58,"N":11554,"t":"Forecasts many feet wearing down ground, writing trample before the verb","l":"This neuron predicts that many feet are about to press on and wear down a surface, and it pushes ' tr', ' tread' and 'traffic' before the verb arrives. Crowds, hikers and herds drive it hardest. Delivery trucks count somewhat less, a single walker much less, and rain, wind or…","st":"tested","i":8,"c":7,"m":[["hikers",102],["crowd",95.5],["cattle",84.0],["vehicles",73.0],["one walker",27.1],["rain",5.0],["nonsense agent",4.6],["wind",3.8],["tree roots",1.4],["sprinkler",-6.1]]},{"id":"L58_N12813","k":"d","L":58,"N":12813,"t":"An 'Emp-' detector amplified by theory-to-evidence turns","l":"This neuron fires on the first piece of the word 'empirical', and only strongly when research writing is turning from theory to evidence. The token alone gives a modest signal. The same token right after 'predictions are not facts' or under a literature-review heading jumps to…","st":"tested","i":5,"c":4,"m":[["theory→data",356],["lit-review heading",288],["paper title",204],["research prose",146],["Kuhn frame",109],["nonsense",38.0],["recipe control",14.9],["Roman",7.5]]},{"id":"L58_N13118","k":"d","L":58,"N":13118,"t":"A neuron that predicts the next kaomoji face","l":"This neuron predicts a text face. It fires on the token just before an eye-first kaomoji such as (^_^) or (T_T) is about to start, most strongly when earlier faces make another one likely. It reaches about +200 across languages. Western :) and emoji get much less, and the same…","st":"tested","i":8,"c":5,"m":[["o_O face",114],["kaomoji",106],["crying kaomoji",100],["price hedge",79.5],["Western :)",2.8],["emoji",1.8],["no-cue control",0.2],["words in parens",-1.0],["cost item",-11.1],["stranger's name",-25.4]]},{"id":"L58_N13860","k":"d","L":58,"N":13860,"t":"A Swedish-only neuron that tracks grammatical gender and word stems","l":"L58/N13860 fires only in Scandinavian text, most strongly in Swedish. Within that text it carries a sign. Its highest values fall in the middle of inflected words such as på|ver|kar, before the ending appears. Neuter noun phrases are positive. Common-gender phrases are strongly…","st":"tested","i":9,"c":6,"m":[["verb stem",117],["Norwegian neuter",51.0],["Swedish neuter",39.5],["nonce neuter",37.8],["neuter definite",29.0],["German control",8.3],["English control",4.3],["nonce common",-105],["Swedish common",-107],["common definite",-112]]},{"id":"L58_N16103","k":"d","L":58,"N":16103,"t":"The Alice and Peter Pan unit, and its Victorian-canon opposite","l":"This unit tracks one thing: the text is inside Lewis Carroll's Alice books or J. M. Barrie's Peter Pan. Any cue is enough, whether a title, a character or an unnamed plot motif. The unit then pushes the model to write names like Alice, Carroll, Peter or Wonderland next. Other…","st":"tested","i":7,"c":8,"m":[["Neverland",212],["Wonderland",188],["Kensington (Barrie)",132],["Oz",74.0],["Narnia",72.5],["Pooh",70.0],["Hogwarts (modern)",32.0],["no story",17.9],["nonsense",7.1],["Wessex (Hardy)",-34.0]]},{"id":"L58_N16829","k":"d","L":58,"N":16829,"t":"A textbook-voice detector that peaks just before the lesson's payload","l":"This neuron fires when text sounds like course material, and hardest when the subject is economics. It peaks not on the subject word but on the connective that promises an explanation, such as 'because' or 'defined as', before that explanation arrives. Question framing lifts any…","st":"tested","i":6,"c":6,"m":[["macro econ",15.6],["econ core",14.9],["accounting",14.1],["nonsense control",10.1],["chemistry",8.2],["CS (low)",5.6],["law of demand",4.8],["news control",3.2],["calculus",2.5],["name control",-11.2]]},{"id":"L58_N18432","k":"d","L":58,"N":18432,"t":"Spanking positive, whip negative: one axis whose negative pole writes 'whip'","l":"This late-layer unit is a signed axis inside corporal punishment. It runs positive whenever a text is about spanking, whether a story, a roleplay request or child-development research. It runs negative, about three times harder, whenever a whip or lash is the natural next…","st":"tested","i":7,"c":9,"m":[["spanking core",74.5],["timeout halo",9.3],["no-cue control",-0.5],["neutral control",-1.0],["school paddling",-28.5],["whisk homonym",-47.5],["riding crop",-87.0],["BDSM spanking",-118],["galley overseer",-241],["coachman reins",-306]]},{"id":"L58_N19593","k":"d","L":58,"N":19593,"t":"A neuron that predicts shouting, and turns negative for hushed speech","l":"This neuron predicts shouting. It fires on the token just before a word like shout, yell or scream is due, in any language. The setup can be an emergency, a deaf listener, noise to shout over, or the rule that all caps means shouting. It writes those words directly, and removing…","st":"tested","i":8,"c":5,"m":[["rock concert",108],["deaf listener",50.5],["fire emergency",50.2],["stadium crowd",47.8],["nonsense",18.0],["neutral place",-3.3],["quiet library",-8.1],["lullaby",-9.7],["animal roar",-18.9]]},{"id":"L58_N1965","k":"d","L":58,"N":1965,"t":"A signed topic-slot neuron: igneous rock positive, Southwest turquoise negative","l":"This neuron fires on the token before a topic word is due and encodes which topic the model expects. Prose about igneous rock bodies and their minerals drives it hardest, with Eastern Woodlands archaeology close behind. The same slot goes negative for Southwest turquoise…","st":"tested","i":3,"c":2,"m":[["igneous sill (core)",61.2],["Iron Age broch",48.0],["Eastern Woodlands",31.0],["basalt flow",29.4],["no-place control",11.4],["Southwest pueblo",9.8],["chalk (sedimentary)",4.5],["Florida control",2.7],["mantle process",-10.0],["Arizona turquoise",-19.9]]},{"id":"L58_N22892","k":"d","L":58,"N":22892,"t":"One axis at the disorder slot: clotting positive, immunity negative","l":"This neuron tracks which body system a named disorder has failed. Wherever a disease is about to be named or has just been named, it reads positive for failures of clotting, from hemophilia to fragile vessel walls, and negative for failures of immunity, deepest for antibody…","st":"tested","i":7,"c":6,"m":[["bleeding (core)",66.0],["factor deficiency",57.8],["vessel wall",50.0],["clot risk",26.1],["neuro control",26.0],["anemia halo",19.0],["nonsense control",8.9],["hormone (neg)",-76.5],["immune (neg)",-172],["antibody (neg pole)",-174]]},{"id":"L58_N22960","k":"d","L":58,"N":22960,"t":"A neuron that anticipates the Holy Grail before it is named","l":"This neuron fires when the Holy Grail, either the legendary cup or the 'holy grail of X' idiom, is about to be named. It peaks on the lead-in ('quest for the', 'holy', 'gr') and drops once the word is complete. It tracks the concept, not the string: the company Grail barely…","st":"tested","i":8,"c":7,"m":[["the word Grail",318],["Grail company",310],["sacred vessel",306],["Christ's cup",292],["Golden Fleece",262],["Excalibur",244],["Ark",244],["nonsense",244],["mundane control",242],["coveted coin",155]]},{"id":"L58_N24387","k":"d","L":58,"N":24387,"t":"A neuron for playful secrets makers hide in software and games","l":"This neuron tracks the easter-egg idea: a playful secret that a maker deliberately hides inside software, a game or other media for users to find. The word itself is not needed. It fires hardest on the small word just before the hidden thing is named, reaching about +220. Famous…","st":"tested","i":8,"c":6,"m":[["Google trick",106],["gaming lore",98.5],["joke in app",79.5],["game in software",61.8],["crafted: novel",50.2],["crafted: building",42.2],["malicious",30.4],["physical hiding",23.1],["nonsense",10.9],["no hiding",3.6]]},{"id":"L58_N24649","k":"d","L":58,"N":24649,"t":"A neuron that predicts the tip for your waiter","l":"This neuron marks the spot where a tip for a service worker is expected. It fires hardest on the token just before the tip word, such as 'left a generous' or 'a 25%', and it writes that word: zeroing it cuts P(' tip') from about 0.9 to under 0.1. It does not simply track the…","st":"tested","i":8,"c":7,"m":[["waiter",254],["valet",254],["bartender",236],["housekeeper",207],["official (bribe)",176],["bookie",165],["nonsense",134],["landlord (fee)",127],["relative",120],["institution",106]]},{"id":"L58_N2719","k":"d","L":58,"N":2719,"t":"An Ashlag Kabbalah idiom neuron that also predicts against 'divine'","l":"This neuron stays on throughout writing in the Ashlag/Laitman Kabbalah idiom: desire to receive, bestowal, equivalence of form, the light of correction. It peaks on the word just before the next doctrinal term. The idiom, not the topic, sets it off, so a management seminar using…","st":"tested","i":6,"c":7,"m":[["gematria",92.0],["no frame",88.0],["Zohar",85.0],["Ashlag core",84.0],["Hermetic Sefirot",80.5],["Sufi/Creator",75.5],["Christian mystic",72.5],["Chassidic",70.0],["management frame",47.8],["literal secular",14.9]]},{"id":"L58_N3615","k":"d","L":58,"N":3615,"t":"A taste axis running from sappy romance to critics' darlings","l":"This neuron reads the taste tier of a film, book, show or song. It turns positive for sentimental mass-audience romance, such as rom-coms, Hallmark movies and love ballads, and peaks when the work is also judged sappy or clichéd, at the slot where the next title or genre is due.…","st":"tested","i":9,"c":6,"m":[["Hallmark fans",124],["romance readers",105],["romantics",89.5],["soccer moms",81.0],["action",18.5],["mainstream",16.4],["non-media control",-15.0],["film students",-32.8],["arthouse",-44.8],["Pitchfork",-57.2]]},{"id":"L58_N3822","k":"d","L":58,"N":3822,"t":"A next-token dial that anticipates all-, multi- and general-purpose","l":"This neuron works mostly through its negative side. It swings down at the moment the model is about to write a versatility compound such as all-purpose, multi-purpose or general-purpose, scaled by how likely that word is, and ablating it measurably weakens the compound…","st":"tested","i":5,"c":8,"m":[["kernel island",5.6],["nonsense control",-2.0],["low-prior control",-3.7],["multi- non-versatile",-9.3],["Python (core)",-20.5],["keyboard (positive)",-20.9],["compound itself",-24.6],["general- halo",-35.8],["versatile object",-39.2],["versatility idiom",-55.5]]},{"id":"L58_N417","k":"d","L":58,"N":417,"t":"A signed voice axis: Victorian British versus American vernacular","l":"This neuron reports which voice a passage is in, and it fires at the next open slot rather than on the cue words. A British Victorian voice (Holmes, cabmen, shillings and pence) drives it positive. An American vernacular voice (cowboys, baseball talk, folksy letters) drives it…","st":"tested","i":7,"c":5,"m":[["Watson",36.8],["costermonger",19.8],["Yorkshire farmer",8.8],["formal American",6.5],["modern Briton",5.9],["no speaker",5.0],["Obama",-0.9],["cowboy",-9.5],["Kansas farmer",-10.1],["Texas oilman",-13.6]]},{"id":"L58_N6728","k":"d","L":58,"N":6728,"t":"Expecting 'gaps': positive for generic gaps, negative for one specific fault","l":"This neuron fires on the word just before a noun for a break in something whole: gaps, pauses, duplicates, missing values. The noun itself is silent; 'with no', 'without', 'fill these', 'during natural' carry the firing, and plural 'there are' is positive. When one specific…","st":"tested","i":5,"c":6,"m":[["without",56.2],["many",38.2],["there are several",25.9],["there are (plural)",25.9],["has no",25.2],["exactly two",22.5],["off-class control",-5.6],["neutral control",-5.6],["there is a (singular)",-18.0],["there's (singular)",-23.1]]},{"id":"L58_N7809","k":"d","L":58,"N":7809,"t":"Album artwork slots versus film stills and photo credits","l":"This neuron tracks where a picture is going or where it came from. It fires positive at the slot where artwork for a music release is about to be named, such as a CD booklet's inside back cover or an MP3 file's embedded-cover field, peaking near +98. It fires negative when…","st":"tested","i":5,"c":6,"m":[["CD single",67.0],["vinyl LP",66.0],["music, no art",64.5],["album",56.8],["film material",54.8],["DVD",54.2],["nonsense",49.2],["game box",48.8],["no product",42.2],["cookbook",40.5]]},{"id":"L58_N8295","k":"d","L":58,"N":8295,"t":"The neuron that fires just before 'alibi' and writes it","l":"This neuron fires at the moment a text is about to name the account that clears a suspect, the word just before 'alibi'. It does this in murder mysteries, true crime and interrogations across several languages. Everyday excuses and 'airtight' modifiers get a weaker echo.…","st":"tested","i":8,"c":8,"m":[["murder suspect",163],["robbery suspect",162],["witness vouches",128],["boss excuse",64.5],["missed exam",56.5],["neutral control",40.0],["civil legal",34.5],["dinner escape",30.8],["nonsense",23.5],["forensic tools",-39.8]]},{"id":"L58_N8880","k":"d","L":58,"N":8880,"t":"Egyptian gods described (+), hurt or sorry people (−)","l":"This neuron tracks what a sentence is saying about someone. When the subject is an Egyptian god, it switches on and peaks on the word that opens the next epithet, role or cult site, such as 'lord[ of]'. Greek and other gods get only a faint echo. When the subject is a person who…","st":"tested","i":5,"c":8,"m":[["Khnum (core)",46.5],["Ptah",37.5],["Amun-Ra",35.0],["pharaoh",31.2],["secular Egypt",28.2],["Canaanite god",6.8],["no deity",4.7],["Greek god",3.3],["harmed person",-3.7],["Sumerian god",-15.3]]},{"id":"L59_N1130","k":"d","L":59,"N":1130,"t":"An 'interaction' predictor that fires just before the word appears","l":"This neuron carries one signed axis about the concept 'interaction'. Its working pole is negative: it fires -100 to -290 when the model is about to produce an interact- word, in any script. The largest values sit on the stub ' inter' before…","st":"tested","i":6,"c":5,"m":[["interesting control",-3.1],["interface likely",-30.1],["bare stub",-34.5],["international control",-35.2],["nonsense control",-40.8],["interface competes",-108],["protein-protein",-148],["physics interaction",-157],["statistical interaction",-185],["drug-drug (core)",-197]]},{"id":"L59_N11485","k":"d","L":59,"N":11485,"t":"Writes legal-drafting 'such' after shall-verbs and continuing prepositions","l":"This neuron predicts the drafting word 'such'. It fires at the open slot in legal or governance text where boilerplate says 'such X as', for example after an office-holder 'shall perform' or after 'or at'. Turning it off drops P(such) from 0.91 to 0.01. Weaker echoes appear in…","st":"tested","i":8,"c":7,"m":[["shall + transitive",205],["will",173],["'or at' clause",164],["may + transitive",124],["statute until",27.6],["intransitive",3.5],["no frame",3.1],["past tense",-2.2],["name fragment",-19.4]]},{"id":"L59_N12449","k":"d","L":59,"N":12449,"t":"Just before someone names the cozy feeling","l":"Fires just before a word for coziness gets said: “Hygge is a Danish word for”, “…make the living room feel”, or a candlelit snowy scene followed by “The Danes have a word for this:” (+260). Hygge, koselig, gemütlich, gezellig and Korean 아늑하다 all count. Oddly, zeroing it makes “…","st":"tested","h":"../notes/L59_N12449/","i":8,"c":7,"m":[["a word for this:",260],["Hygge –",199],["Danish word for",151],["koselig",147],["living room feel",128],["cough from a cold",86.5],["Schadenfreude",0.83],["meditation apps",-38.8]]},{"id":"L59_N1251","k":"d","L":59,"N":1251,"t":"A signed slot that predicts deep-sea versus inshore animals","l":"This neuron fires on the token just before the next sea animal is named, such as a comma, 'and', or the article 'a', and its sign predicts which kind of animal comes next. After deep-sea or midwater animals such as lanternfish and anglerfish it goes strongly positive. After…","st":"tested","i":7,"c":7,"m":[["midwater fauna",104],["deep fishery",47.5],["hadal fish",40.2],["no animal",9.4],["freshwater",-9.4],["open ocean",-16.2],["generic abyss",-24.4],["reef",-29.6],["regional sea",-73.5],["inshore catch",-80.0]]},{"id":"L59_N13688","k":"d","L":59,"N":13688,"t":"A neuron that anticipates 'möglich' completing 'as fast as possible'","l":"This neuron predicts that 'possible' is about to complete a phrase of maximal feasibility. It fires hardest inside German compounds such as 'schnellstmöglich' and 'bestmöglich', on the piece just before '-öglich' (up to about +600). It fires roughly three times weaker before a…","st":"tested","i":6,"c":5,"m":[["schnell|stm",356],["best|m",286],["open stem",233],["alle| mö",232],["English as…as",209],["so … wie",190],["French phrase",150],["no-möglich control",35.0],["'maybe' control",0.0]]},{"id":"L59_N15363","k":"d","L":59,"N":15363,"t":"About to say 'crook of your elbow'; beer lists drive it negative","l":"This neuron fires at the point where text is about to name the bend of a human arm: ' cro' before 'crook of your elbow', or 'your' before 'elbows'. It fires hardest in cough-and-sneeze hygiene advice, above +100. Other joints and other meanings of 'crook' fire only weakly. Its…","st":"tested","i":6,"c":7,"m":[["hygiene + handwash",98.0],["arm vein",67.0],["bare cough",61.8],["baby in arm",56.0],["knee",55.0],["criminal",25.8],["river bend",25.0],["crocodile",16.8],["no cue",13.8],["beer context",-61.8]]},{"id":"L59_N18135","k":"d","L":59,"N":18135,"t":"Nets, pots and snares coming next versus who found the shipwreck","l":"This neuron predicts that capture gear comes next. It fires hardest just before words like gillnets, crab pots, longlines or snares, especially when whales, turtles or gorillas are at risk. It also responds more weakly to trap-like hazards such as landmines and fouled…","st":"tested","i":7,"c":7,"m":[["right whales",176],["gorillas (snares)",173],["sea turtles",125],["no-gear wildlife",114],["vaquita",104],["propeller",101],["trees (vines)",16.6],["nonsense",5.9],["shipwreck",-43.2]]},{"id":"L59_N18234","k":"d","L":59,"N":18234,"t":"A poison-frog detector with a reef-fish halo and a yarn pole","l":"This neuron tracks whether the animal under discussion is, or is about to be, a small tropical frog, and poison dart frogs top the scale. It fires on the species name and also on the slot before it when context predicts one; 'a toxic, brightly colored' reaches about +120 before…","st":"tested","i":6,"c":6,"m":[["Oophaga",91.0],["dart frog",62.2],["common frog",37.0],["reef fish",24.2],["glass frog",17.5],["toxic snail",11.2],["bullfrog",4.7],["no animal",-0.4],["colourful control",-2.5],["yarn",-4.0]]},{"id":"L59_N18389","k":"d","L":59,"N":18389,"t":"A slot detector for Agatha Christie in catalogs of detective fiction","l":"This neuron marks the moment just before a classic detective-fiction name is written, as in the 'by' after a Christie title or the 'created by' after Poirot. Agatha Christie is the strongest case. Other golden-age detectives fire progressively less. The same slot goes negative…","st":"tested","i":7,"c":6,"m":[["Christie title",179],["Poirot",167],["Holmes",116],["Austen",62.5],["Maigret",28.4],["no-book control",10.8],["modern thriller",-6.8],["Nabokov",-13.6],["Dan Brown",-28.2],["Rebus",-59.0]]},{"id":"L59_N19626","k":"d","L":59,"N":19626,"t":"A leading-digit vote for numbers the model is about to compute","l":"This neuron sits in the slot just before a number the model is working out from context: a percentage, a conversion, a product, a hex digit. It ignores the topic and reads the leading digits of the model's own upcoming answer. Values starting 55 to 59 drive it strongly positive…","st":"tested","i":10,"c":7,"m":[["→ 573 (core, x10)",90.0],["→ 18 (side band)",52.5],["→ 57 (core)",40.5],["→ 5.7 (core, /10)",17.5],["→ 29 (side band)",8.8],["→ 11 (side band)",4.0],["→ 62 (negative)",2.7],["no value (control)",-0.3],["→ 50 (minor dip)",-8.2],["→ 31 (negative)",-15.4]]},{"id":"L59_N21520","k":"d","L":59,"N":21520,"t":"A neuron that predicts 'place' in 'place an order'","l":"This neuron predicts the verb 'place' in 'place an order'. It fires on the token just before a customer's ordering verb is due, such as 'the date you ___' in a shop email, and it does this across languages. Ablating it drops 'placed' from 0.97 to 0.06, and 'ordered' or 'made'…","st":"tested","i":7,"c":8,"m":[["betting",115],["shop thanks",75.0],["refund/order",62.2],["booking",29.9],["broker",19.0],["bank service",18.5],["nonsense",4.6],["cemetery",-3.8],["genealogy",-3.9],["museum",-9.1]]},{"id":"L59_N22621","k":"d","L":59,"N":22621,"t":"Herod is next (+) versus Persian capital is next (−)","l":"This neuron fires on the word just before a name is due in ancient Near Eastern narrative, and its sign says which kind of name. It goes strongly positive (around +170) when a ruler of Gospel-era Judea is next, usually Herod but also Pilate or Tiberius, and it pushes the model…","st":"tested","i":7,"c":6,"m":[["Galilee / Baptist",235],["Roman Judea",144],["Persian Susa",107],["Maccabean",99.5],["Rome",93.5],["Babylon",89.5],["Athens",49.5],["modern city",40.8],["Egypt",20.9],["nonsense",20.2]]},{"id":"L59_N2396","k":"d","L":59,"N":2396,"t":"A geothermal-place neuron: geysers positive, lava tubes and reservoirs negative","l":"This neuron tracks whether the text is about a hydrothermal place such as Yellowstone, Rotorua or a geyser field. It fires most at open slots inside that topic, such as list commas and the start of feature words, rather than on the place name. How strongly it fires depends on…","st":"tested","i":4,"c":8,"m":[["Yellowstone",191],["Rotorua",165],["Geysir",153],["spa town",121],["icy moon",80.0],["cold spring",17.2],["no place",1.8],["non-thermal park",-7.8],["reservoir",-28.4],["lava tube",-51.2]]},{"id":"L59_N4479","k":"d","L":59,"N":4479,"t":"A tree-ring evidence neuron that peaks where proof of the past goes","l":"This neuron tracks one idea: tree rings serving as evidence about the past. That covers climate reconstructed from ring widths and wooden objects dated by their rings. It fires in any language, stays on across a passage, and peaks where ring evidence is the expected next idea.…","st":"tested","i":6,"c":4,"m":[["pine stand",72.0],["drought (core)",59.5],["ice-core halo",47.8],["oak panel (core)",34.0],["otolith halo",34.0],["non-climate control",17.4],["wood, no past",14.4],["nonsense",14.4],["space (negative)",-33.2]]},{"id":"L59_N7406","k":"d","L":59,"N":7406,"t":"Signed topic readout: panels and balers up, patents and title law down","l":"L59/N7406 reads out a handful of late-layer topic detectors with a sign attached. It fires positive across whole passages about photovoltaic module layer stacks and baling equipment, and more moderately across contract and civil-law doctrine. It fires negative across…","st":"tested","i":2,"c":2,"m":[["PV layers (core)",65.0],["baler (core)",23.0],["shredder (halo)",9.9],["forklift (format only)",1.5],["nonsense control",-0.5],["control",-1.1],["tort/offences (neg)",-4.5],["contract law (tier two)",-12.5],["patent assembly (neg)",-26.2],["deed covenants (neg)",-58.5]]},{"id":"L59_N8093","k":"d","L":59,"N":8093,"t":"A neuron that tracks damp forest cover versus dry open ground","l":"This neuron tracks what kind of ground a wild scene is set on, and fires at the word where the next place is about to be named: a preposition, a list comma, 'such as'. Damp, sheltering forest ground like leaf litter, rotting logs, moss and buttress roots drives it strongly…","st":"tested","i":5,"c":5,"m":[["moss, roots",79.0],["swamp streams",72.5],["rotting logs",55.0],["leaf litter",49.5],["caves, rocks",27.6],["no place",12.0],["garden pots",-18.6],["roadsides",-19.9],["wet open",-26.8],["prairie",-74.5]]},{"id":"L59_N8629","k":"d","L":59,"N":8629,"t":"A neuron that predicts 'change' as transformation, one token early","l":"This neuron fires on the word just before the model expects 'change' in its uncountable sense of transformation: 'agents of', 'the only constant is', 'continuity and'. It does not simply track how likely 'change' is: 'global climate' barely moves it. Its output pushes the model…","st":"tested","i":6,"c":5,"m":[["agents of",102],["rapid",89.0],["drivers of",83.0],["synonym slot",75.0],["climate control",44.8],["resistant to",42.8],["weak process",24.0],["no cue",21.1],["plural changes",4.6]]},{"id":"L59_N925","k":"d","L":59,"N":925,"t":"Favoured jargon runs versus small British towns awaiting their county","l":"This neuron is a signed axis. On the positive side it fires inside runs of a few favoured specialist vocabularies: crab and spider appendages, excavated-site features, phytochemicals, iron minerals, basal-ganglia abbreviations and triangle geometry. Most other jargon leaves it…","st":"tested","i":6,"c":2,"m":[["no place",-2.0],["continental",-7.8],["US town",-14.2],["big city",-19.5],["northern English",-21.4],["south coast",-32.0],["home county",-55.2],["Scottish town",-59.2],["Thames Valley",-62.2],["Welsh town",-63.8]]},{"id":"L5_N13200","k":"d","L":5,"N":13200,"t":"Little wonder, no surprise: the stock phrases for an unsurprising result","l":"An early neuron that fires on “little wonder”, “small wonder”, “no wonder”, and a bit less on “no surprise” and “unsurprisingly”: the phrases for saying the last fact explains the next. It knows the phrases, not the move. “Naturally”, “of course” and “as expected” leave it at…","st":"tested","h":"../notes/L5_N13200/","i":4,"c":7,"m":[["Little wonder",27.1],["Little surprise",21.0],["No wonder",18.8],["Unsurprisingly",15.1],["no surprise",11.2],["puppy: a little wonder",11.0],["Of course",0.05],["off-grid",-27.6]]},{"id":"L60_N10565","k":"d","L":60,"N":10565,"t":"Norwegian context gate, signed by common word versus name or number piece","l":"This neuron multiplies two signals. A gate tracks how Norwegian the surrounding text is, with Basque, Icelandic and Finnish as weaker echoes and Danish and Swedish surprisingly low. The sign says what kind of token comes next: positive on ordinary words and endings in running…","st":"tested","i":5,"c":5,"m":[["French control",-0.3],["no place",-0.3],["Swedish neighbour",-7.9],["Basque halo",-9.4],["Danish neighbour",-12.8],["the word Norway",-15.3],["Icelandic halo",-16.1],["Finnish Arctic",-45.0],["Norwegian region",-66.5],["Norwegian city",-91.0]]},{"id":"L60_N11971","k":"d","L":60,"N":11971,"t":"A baseball dial: home runs push it up, everything else down","l":"This neuron is a signed dial that only moves inside baseball text. It turns positive when home runs are the subject, such as career totals, slugger names and record chases, and it peaks just before 'home run' is said. It turns negative for every other on-field event, including…","st":"tested","i":9,"c":7,"m":[["HR record",266],["HR king",147],["HR relabel",128],["hits record",57.8],["non-sport",40.8],["no cue",40.0],["pitcher",31.6],["cricket",30.5],["glove man",-43.2],["contact hitter",-67.0]]},{"id":"L60_N13449","k":"d","L":60,"N":13449,"t":"A catch-verb-is-coming neuron that peaks mid-word on split verbs","l":"This neuron tracks one thing: a catch or capture verb is about to be written. It rises on the words that set up the catch and peaks on the first fragment of the verb when the verb is split into pieces, as in Russian ' п|оймать' or German 'einz|uf|angen'. It then drops once the…","st":"tested","i":8,"c":8,"m":[["ball",253],["fish",217],["mood / moment",91.5],["butterfly",77.0],["loose bird",66.0],["thief",59.5],["take control",0.9],["buy control",0.8],["code error",-0.4],["nonsense",-1.7]]},{"id":"L60_N13931","k":"d","L":60,"N":13931,"t":"Health provider slot: bone, brain and eye negative, access and coverage positive","l":"This neuron watches health text and fires on the function word right before a specialty, provider or institution is named. Its sign depends on the kind of care coming next. Spine, bone, brain and eye care drive it strongly negative, down to about -130, while soft-tissue…","st":"tested","i":6,"c":6,"m":[["cosmetic (+)",39.8],["ENT/voice (+)",38.5],["sexual health (+)",35.2],["generic medical",18.6],["cardiac control",13.2],["non-medical control",13.2],["eye care",-21.4],["neurosurgery",-35.0],["spine (core)",-49.0],["orthopaedics",-65.0]]},{"id":"L60_N14726","k":"d","L":60,"N":14726,"t":"Candy brand due (+) or fairy-tale figure due (−): a signed name predictor","l":"This late-layer neuron predicts what kind of name comes next. It fires positive at the token right before a chocolate or candy brand is due, and idles at half height anywhere the text is about sweets. It fires negative when a fairy-tale figure is due, from Aladdin to Rapunzel.…","st":"tested","i":7,"c":8,"m":[["candy",134],["chocolate",126],["jewelry",43.5],["ice cream",26.0],["no cue",9.5],["soda",5.6],["cars",-0.4],["novels",-0.9],["fairy tale",-13.3],["Disney",-14.5]]},{"id":"L60_N14825","k":"d","L":60,"N":14825,"t":"Expecting mourning: a neuron that signs the coming verb","l":"This late-layer neuron reads the slot where the next word is chosen and reports what kind of word is coming. When a crowd has assembled and the natural continuation is to honour or mourn someone, it fires strongly positive, peaking around +70 on the word 'to'. When the same slot…","st":"tested","i":6,"c":3,"m":[["mourners",34.2],["all ranks",25.6],["bereaved",25.0],["gratitude",18.9],["pilgrims",17.6],["curiosity",9.1],["engineers",0.8],["institution",-8.0],["no crowd",-8.7],["population",-12.1]]},{"id":"L60_N17524","k":"d","L":60,"N":17524,"t":"Copied digits inside a recognized identifier scheme; negative before span quantities","l":"The neuron fires when the model copies a number it has already seen and that number sits in the arbitrary body of an identifier whose scheme it recognizes: a publisher's ISBN block, a journal's DOI stem, NBER's prefix. Copy alone or scheme alone gives a fraction; together they…","st":"tested","i":7,"c":7,"m":[["copy x real stem",199],["mismatch x real stem",183],["copy, prose",61.5],["copy x other journal",48.2],["copy x PLoS stem",47.2],["copy x fake stem",37.2],["copy x ISBN block",25.0],["copy, no scheme",21.1],["nonsense control",21.1],["no copy x real stem",15.9]]},{"id":"L60_N21681","k":"d","L":60,"N":21681,"t":"A neuron that writes the start of an expected 'hang' word","l":"This neuron fires wherever the model expects a 'hang' word next, whether or not one follows, and helps write its opening piece. 'A coat on a hook still' triggers it; 'a pair of boots still' does not. It covers phone hang-ups, pictures on walls, laundry, overhangs and low-hanging…","st":"tested","i":8,"c":9,"m":[["drying garland",364],["gallows",268],["portrait",162],["coat on hook",160],["overhang",17.9],["Hung- lookalike",10.6],["boots (sit)",4.4],["stack (lie)",-0.6],["nonsense",-1.1],["no cue",-1.2]]},{"id":"L60_N21686","k":"d","L":60,"N":21686,"t":"Formal Vietnamese syllables up; California-coast place slots down","l":"This neuron has no single concept. Its strongest positive firing is on the first syllable of formal Sino-Vietnamese words such as địa or Công, with smaller positive bumps from a basketball word-problem stem and reference-style prose. Its negative pole is the word just before a…","st":"tested","i":4,"c":3,"m":[["no place",0.9],["Kyoto",-2.9],["Texas control",-10.6],["Cape Cod",-23.9],["SoCal city",-30.6],["Oregon",-31.0],["California",-33.5],["Marin",-47.5],["Big Sur",-50.0],["Point Reyes",-51.8]]},{"id":"L60_N22330","k":"d","L":60,"N":22330,"t":"A neuron that predicts 'they were unprepared' before it is said","l":"This neuron fires on the token just before a text says that someone facing a crisis or demand was unprepared, caught off guard or not equipped. It needs both the demand and a cue that the verdict will be negative, such as 'admitted' or 'despite drills'. The peak comes before the…","st":"tested","i":8,"c":6,"m":[["admitted",195],["despite drills",114],["without warning",71.5],["neutral 'said'",55.5],["boasted",46.2],["thanks to drills",15.4],["other predicate",0.7],["nonsense",-2.2],["labor list",-5.0]]},{"id":"L60_N25269","k":"d","L":60,"N":25269,"t":"A take-versus-bring switch for verbs of carrying things somewhere","l":"This late-layer neuron decides direction for verbs of carrying. Right before the model says such a verb, mostly in Portuguese, Spanish, Russian or Hungarian, it fires and pushes take/levar/llevar over bring/trazer/traer. Turning it off flips 'levar' to 'trazer'. In English the…","st":"tested","i":9,"c":8,"m":[["go to Ana's",256],["hospital visit",172],["trip packing",165],["pet to vet",136],["no conveyance",126],["nonsense",79.0],["come here",70.0],["upbringing (PT)",29.6]]},{"id":"L60_N3251","k":"d","L":60,"N":3251,"t":"WoW's next-term slot, shared with trading, inverted by audio software","l":"This neuron fires on the small word just before a specialized term is due, and the surrounding topic sets the sign. World of Warcraft prose drives it hardest, reaching about +70 on the slot before 'character screen', with other online games following in order of how WoW-like…","st":"tested","i":4,"c":4,"m":[["WoW vocabulary",61.2],["WoW name",58.2],["trading platform",41.0],["other MMO",30.8],["non-MMO game",18.6],["nonsense",18.4],["generic software",11.7],["no-domain control",8.0],["IBM product",-22.1],["audio software",-28.4]]},{"id":"L60_N5835","k":"d","L":60,"N":5835,"t":"A neuron that pushes 'cat' next and flickers on meow","l":"This neuron has two poles. It goes strongly negative on the token just before the model writes 'cat', either the animal (cat, chat, Katze) or a 'cat'-initial piece such as categories or cater. It goes positive on ' Me' when a cat is about to say 'meow'. Only the negative pole…","st":"tested","i":9,"c":6,"m":[["word 'cat'",8.8],["wolf prey",1.6],["dog (contrast)",1.4],["big-cat prey",1.1],["nonsense",0.8],["human control",0.4],["bird prey",-6.6],["fish prey",-13.6],["mouse (core)",-39.5]]},{"id":"L60_N6946","k":"d","L":60,"N":6946,"t":"Memorized name completions up, ranking and planting slots down","l":"This neuron is not a drought detector. It fires on the token where a specific memorized name locks in. The clearest case is 'Palmer D|rought' Severity Index. Agency names such as 'and Arnold' Worldwide also trigger it. Near-miss names and plain drought prose stay quiet. Its…","st":"tested","i":3,"c":4,"m":[["Palmer Drought",32.2],["other instrument",17.9],["sister index",14.8],["leading token",13.8],["drought agency",11.1],["plain prose",5.6],["other memorized name",4.9],["fake index",2.2],["plain institution",0.1],["near-miss name",-0.2]]},{"id":"L60_N7410","k":"d","L":60,"N":7410,"t":"A neuron that recognizes the given value of a memorized exercise","l":"This neuron fires on the number in a math or coding exercise that confirms the model is reading a problem it has memorized, for example the 1 in 'BC=1' after a stored cosine given. Change one word or value of the stored wording and it goes silent. Values that come after the…","st":"tested","i":9,"c":6,"m":[["nonsense control",-2.1],["symbolic given",-5.5],["plain fraction",-7.1],["generic value",-7.2],["changed denominator",-8.5],["partial match",-12.2],["stored problem",-12.3],["near variant",-14.1]]},{"id":"L60_N8897","k":"d","L":60,"N":8897,"t":"A spacing dial: will the next math operator arrive glued or spaced?","l":"Inside formulas and code, this neuron predicts how the next operator will attach to the current token. It fires positive when the operator continues without a space: the backslash glue before \\leq, an align ampersand, an HTML entity fragment, a variable in compact 'x∈Z'. It goes…","st":"tested","i":5,"c":8,"m":[["entity '&lt'",33.0],["compact 'b'",23.9],["C compact 'b'",4.4],["no math",0.7],["prose 'b'",-8.1],["C spaced ' b'",-42.2],["align ' &'",-65.5],["glue ' \\'",-70.5],["closed formula '$'",-70.5],["spaced 'b'",-70.5]]},{"id":"L61_N11504","k":"d","L":61,"N":11504,"t":"A sound dial: describing music versus singling out one note","l":"This neuron is a signed dial that moves only when text is about sound, rhythm or pulse. It reads positive on words that describe how a sound is organized (tempo, regularity, intervals, pitch), most strongly in Russian and Polish music answers. It reads negative on a selector…","st":"tested","i":7,"c":6,"m":[["no sound",-48.0],["Morse code",-53.2],["bird",-56.2],["kettle",-67.5],["jet engine",-71.5],["songbird",-74.0],["null control",-76.0],["soprano",-99.5],["organ",-108],["orchestra",-112]]},{"id":"L61_N16502","k":"d","L":61,"N":16502,"t":"A cond- neuron: brakes official condemnation, boosts non-English condensation","l":"This late-layer neuron always writes against cond- tokens, but what triggers it depends on its sign. It fires positive just before governments, the UN or whole communities publicly condemn an atrocity, and there it holds back 'condemned'; ablating it at the top example raises…","st":"tested","i":9,"c":6,"m":[["multi-actor list",117],["Secretary-General",64.5],["UN Council",48.8],["community",41.5],["nonsense",32.2],["opposition",19.9],["neighbors",7.4],["legal/building",0.7],["aunt",0.4]]},{"id":"L61_N18632","k":"d","L":61,"N":18632,"t":"A trade-versus-housekeeping switch that fires where the worker noun comes","l":"This neuron tracks what kind of work a text is about. Craft, guild, union and hired-trade writing pushes it positive, and hotel housekeeping pushes it negative. The state is present throughout such text, but it peaks two- to threefold on the word just before a worker noun is…","st":"tested","i":6,"c":5,"m":[["cab firm",22.9],["bar",16.6],["guild",-13.6],["software",-13.9],["profession",-15.4],["craft union",-17.9],["kitchen",-26.0],["no topic",-31.6],["hotel",-47.5],["housekeeping",-61.5]]},{"id":"L61_N21523","k":"d","L":61,"N":21523,"t":"Marks an -ist person noun being spelled, one or two tokens before its suffix","l":"L61/N21523 fires on a mid-word fragment when the model expects the word under construction to end as a person noun in -ist (also -ista, -iste, -ист). In the top 100 corpus firings the peak sits one token (58) or two tokens (41) before the word ends, on pieces like omet…","st":"tested","i":6,"c":7,"q":[["p0_01",-11.9],["p0_1",-5.11],["p1",-1.02],["p5",-0.503],["p10",-0.363],["p25",-0.158],["p50",-0.016],["p75",0.121],["p90",0.319],["p95",0.503],["p99",1.37],["p99_9",9.3],["p99_99",62.9]]},{"id":"L61_N9049","k":"d","L":61,"N":9049,"t":"Latin pending-verb slots positive, learned definitions and lists negative","l":"This neuron acts as a gated, signed detector. Latin, or learned and doctrinal English, opens the gate. The local slot then sets the sign. It fires positive where a Latin clause is building toward a verb that has not yet arrived, peaking near +120 on infinitives before 'ausus…","st":"tested","i":6,"c":6,"m":[["pending complement",100],["habitual Latin",54.8],["plain English",-0.4],["nonsense",-0.4],["Italian",-1.7],["philological English",-34.0],["Latin title in English",-40.2],["verb already given",-51.8],["grammar drill",-53.8],["Latin alternative",-78.5]]},{"id":"L62_N10233","k":"d","L":62,"N":10233,"t":"A brake on 'respond' when the passage sets it up","l":"This neuron fires on the word just before an expected 'respond' or 'response', and on répondre in French. It then subtracts from that respond-family logit, acting as a brake. Ablating it raises P(respond), and steering it up cuts P(respond) sharply. What drives it is the…","st":"tested","i":8,"c":6,"m":[["MSS tumors",75.0],["low-TMB patients",58.2],["nonsense",34.2],["servers",26.9],["receptors",9.5],["bridges",-7.1],["prices",-11.4],["customers",-11.9],["administrators",-16.8]]},{"id":"L62_N15550","k":"d","L":62,"N":15550,"t":"City word opening a newspaper name, amplified by radical-politics history","l":"This neuron fires on the first word of a newspaper or institution name, usually a city, as in 'the Detroit' Institute or 'the Amsterdam' News. Some openers fire on their own; most rise sharply when the text is about 1930s labor or radical politics, with Seattle climbing from…","st":"tested","i":6,"c":4,"m":[["Daily",50.2],["Detroit",49.2],["Amsterdam",49.0],["Oakland",26.8],["Seattle",26.5],["Chicago",22.4],["Boston",17.0],["no name",6.5],["leftist word",-5.2],["era word",-11.6]]},{"id":"L62_N18552","k":"d","L":62,"N":18552,"t":"Separators that land at the right margin of hard-wrapped text","l":"This neuron fires on commas, periods and similar separators in hard-wrapped plain text, such as OCR'd books, footnotes and reference lists, when the token lands near the right margin set by earlier lines. In other words, it marks the point where a line break is due within a word…","st":"tested","i":10,"c":6,"m":[["60w, col 60",91.0],["60w, col 76",86.5],["80w, col 76",82.0],["80w, col 90",60.8],["60w, col 48",51.2],["80w, col 60",47.0],["60w, col 33",28.2],["no wrap",16.8]]},{"id":"L62_N25219","k":"d","L":62,"N":25219,"t":"Sense-sorting corrector for 'marked' one layer before the output","l":"One layer before the output, this neuron edits the model's bet on the word family mark. It fires on the token before a slot where 'marked' could come next. In the labeling, grading or notable sense ('the socket [marked] COM', 'a [marked] increase') it fires positive and trims an…","st":"tested","i":8,"c":8,"m":[["copy of 'marked'",35.5],["labeled",26.4],["stamped",24.4],["turbulent era",22.8],["labeling cue",3.4],["no cue",1.2],["clinical degree",0.8],["markup",-6.3],["con victim",-6.9],["football marking",-18.0]]},{"id":"L62_N2529","k":"d","L":62,"N":2529,"t":"A two-pole aequ- neuron: prose 'equaled' up, pre-equator slots down","l":"This neuron tracks the Latin aequ- word family, with two poles at different positions. It fires positively on the fragment ' equ' when prose will finish it as 'equaled', as in prices that equaled a figure or composers no one has equaled. It fires negatively on the word just…","st":"tested","i":6,"c":6,"m":[["jobless rate",95.5],["bread price",67.5],["currency",58.0],["donation amount",48.8],["register value",47.5],["sprinter time",36.5],["nonsense",19.2],["donation sum",14.0],["angle sum",8.5]]},{"id":"L62_N3407","k":"d","L":62,"N":3407,"t":"Word-onset neuron: open verb onsets positive, locked completions negative","l":"This late-layer neuron fires at the start of a word the model is about to finish. It is positive where the completion is still open: bare-verb openers like 'to', 'and' and modals, and above all the ' we' of 'weep'. It is negative on onsets whose ending is locked in, such as…","st":"tested","i":3,"c":6,"m":[["we|ep",48.8],["noun control",43.5],["st| onset",21.4],["bare 'to'",18.2],["slot opener",14.9],["rev| onset",13.5],["mis| locked",-3.6],["bel|ittle",-9.4],["wh|ine",-41.5]]},{"id":"L62_N4494","k":"d","L":62,"N":4494,"t":"A dial for expected ion, particle and protein words","l":"This neuron tracks what the model expects next in science text. It fires strongly negative when context makes an -ion/-on particle word or 'protein' nearly certain, so the same stem ' ferm' reads about -86 before fermions and 0 before fermentation. Its positive side lights up in…","st":"tested","i":6,"c":5,"m":[["yeast control",7.5],["brewing control",-32.2],["virology",-38.2],["redox chain",-44.5],["spin statistics",-44.8],["nonsense",-45.0],["proteins",-65.0],["atomic particles",-65.5],["non-science",-65.5],["quarks/leptons",-74.5]]},{"id":"L62_N4621","k":"d","L":62,"N":4621,"t":"A subject-awaiting-verb detector for humanistic argument","l":"This neuron fires on the subject of a clause, the noun, pronoun or 'which/who' that is still waiting for its verb, in long English argument about literature, faith, nations or values. The signal builds as the paragraph grows more argumentative. Hyphenated URL slugs and math…","st":"tested","i":4,"c":4,"m":[["author name",35.5],["science halo",33.0],["relative onset",27.4],["literary subject",20.0],["nonsense",8.0],["theology subject",5.5],["policy subject",-10.2],["neutral noun",-10.2],["measurement",-20.2],["Spanish procedural",-46.8]]},{"id":"L62_N889","k":"d","L":62,"N":889,"t":"A Dutch-versus-Finnic language gate that also scores events against feelings","l":"L62/N889 multiplies a language gate by a signed content axis. The gate opens on Dutch, Afrikaans and Frisian, and also on Estonian, Finnish and Latvian, but stays shut on English, German and French. Inside the gate, Dutch-family text reads positive and Finnic-Baltic text…","st":"tested","i":9,"c":6,"m":[["Afrikaans",23.8],["Frisian",9.7],["Finnish",3.0],["Hungarian",-0.6],["German control",-0.7],["nonsense control",-0.7],["French control",-0.9],["Dutch (core)",-9.3],["Catalan halo",-10.8],["Estonian (neg)",-14.6]]},{"id":"L63_N13910","k":"d","L":63,"N":13910,"t":"Sparse relay: early 'that', loop sinks, memorized phrases versus Chinese 是-verdicts","l":"This last-layer neuron is near zero almost everywhere. It switches on in a few narrow, sink-like situations: ' that' or ' which' in the first three tokens of a cut-off document, degenerate repetition loops where the repeated token becomes a second attention sink, and a few…","st":"tested","i":3,"c":3,"m":[["no 是",4.6],["English Yes",-0.1],["neutral 是",-3.1],["确实",-7.2],["没错",-8.6],["control opener",-12.9],["你是[错]",-17.1],["opening 是",-30.8],["是的",-30.8],["你是[对] verdict",-56.0]]},{"id":"L63_N15257","k":"d","L":63,"N":15257,"t":"The last-layer neuron that glues '>' to the token before it","l":"This final-layer neuron fires on the token that will be followed by an angle bracket or comparison sign written with no space: a closing tag name before '>', the operand before '>' in an inequality, the digit in a shell '2>'. It detects the coming sign but only decides its…","st":"tested","i":5,"c":8,"m":[["closing <\/code",189],["shell fd 2",127],["include header",107],["LaTeX operand",79.5],["variable, no sign",73.5],["opening <span",71.0],["generic usage",58.8],["regex (?P",32.0],["declared generic",23.6],["prose control",-2.6]]},{"id":"L63_N3186","k":"d","L":63,"N":3186,"t":"A signed dial on the token 'output': brake in templates, boost in prose","l":"This late-layer neuron reads the token just before the model expects to write 'output', in any form (Output, .output, OUTPUT, 输出), and writes almost only onto that token family. Its sign depends on context. In I/O templates, code and instructions it fires positive and slightly…","st":"tested","i":5,"c":7,"m":[["code .output",141],["I/O template",112],["instruction",44.8],["amplifier halo",29.6],["German control",11.4],["no-cue control",1.4],["rated (prose)",-85.5],["turbine shaft",-110],["selector",-144]]},{"id":"L63_N960","k":"d","L":63,"N":960,"t":"Signed spikes at fixed tokens inside a few memorized web strings","l":"This last-layer neuron is silent on almost all text and fires only inside a small set of verbatim, heavily duplicated web strings, spiking to roughly +80 or -90 on one fixed token. Which strings get which sign is arbitrary: the Gaokao 'Elective 4' header and a Slashdot deal…","st":"tested","i":1,"c":2,"m":[["Romanesque (halo)",9.8],["Baroque (core)",6.9],["Rococo (halo)",4.9],["Brutalist",-0.5],["nonsense",-0.5],["Bauhaus (wrong era)",-0.6],["dialect control",-0.7],["Renaissance",-0.8],["disease control",-1.0],["Gothic",-1.4]]},{"id":"L6_N14483","k":"d","L":6,"N":14483,"t":"One gate, two signs: assembling NAM, finishing Land, Way and Manet","l":"This layer-6 neuron reads letters and word pieces, not topics. It fires positive when an all-caps N fragment is completed into NAM or NMA, whatever the surrounding subject. It fires two to three times harder, negative, on three unrelated tokens: 'way' as intensifier or surname…","st":"tested","i":4,"c":4,"m":[["N|AM core",24.2],["lowercase in address",10.3],["whole-token control",1.5],["Monet control",0.8],["other N acronym",0.0],["plain surname control",0.0],["surname Way",-27.0],["surname Land",-48.0],["place Land",-50.5],["Man|et",-63.5]]},{"id":"L6_N18226","k":"d","L":6,"N":18226,"t":"'Table' signed by its modifier: Excel positive, rainbow and water negative","l":"This neuron reads the word 'table' and signs it by the word just before it. The gate opens on any table token, and the modifier sets the sign. After Excel, CSV or another file-format word it fires strongly positive, peaking around +38. Data-sense tables are mildly positive and…","st":"tested","i":3,"c":5,"m":[["Excel",33.8],["CSV",29.4],["Google Sheets",19.4],["HTML",17.6],["no modifier",1.9],["furniture",0.3],["database",0.2],["coffee",-1.5],["water",-16.5],["rainbow",-33.0]]},{"id":"L6_N22538","k":"d","L":6,"N":22538,"t":"HEA, slow burn, insta-love: the labels readers put on a story","l":"An early neuron for the shorthand readers use to label a book: happily ever after (HEA), slow burn, insta-love, #OwnVoices. In a romance review these reach +46 to +60; plain descriptions get +2 to +11. It tells romance HEA from Ireland’s Higher Education Authority, but slow-burn…","st":"tested","h":"../notes/L6_N22538/","i":6,"c":5,"m":[["#OwnVoices",60.0],["insta-love",58.8],["a guaranteed HEA",51.8],["a slow burn",46.8],["slow-burn hot sauce",27.5],["Irish HEA agency",10.8],["plain paraphrase",3.92],["choir's own voices",0.024]]},{"id":"L6_N8982","k":"d","L":6,"N":8982,"t":"A signed list of rare names: ends-here positive, continues negative","l":"This neuron holds a signed, memorized list of rare spelled strings, mostly surnames and place names. It fires strongly positive on the token that finishes names such as Capps, Dubs and Spiers, and strongly negative on look-alike pieces that sit inside longer names, such as…","st":"tested","i":3,"c":7,"m":[["core Capps",34.2],["core Dubs",27.1],["core Spiers",26.2],["off-list Kapps",20.8],["plural noun",5.3],["nonsense",0.7],["common surname",0.2],["neg whole word",-15.4],["neg Gaits",-31.6],["neg Brails",-36.2]]},{"id":"L6_N9349","k":"d","L":6,"N":9349,"t":"An Armageddon word detector that works in every script","l":"This neuron recognizes one word, Armageddon, in any spelling or script. It fires at the token where the word becomes identifiable, such as 'aged' in Arm-aged or 'ед' in Армагеддон. Topic is irrelevant: a care-home company called Armaged scores as high as the apocalypse. Its…","st":"tested","i":5,"c":5,"m":[["Cyrillic",102],["Armageddon",92.5],["company name",92.5],["katakana",76.0],["portmanteau",50.2],["Megiddo",46.0],["'aged' alone",17.6],["other apocalypse",0.0],["doom synonym",-0.4],["belted (neg)",-32.5]]},{"id":"L6_N9376","k":"d","L":6,"N":9376,"t":"Saying it twice: the second “far” in far, far away","l":"Fires on the second copy of a word said twice in a row: the second duck in “Duck, duck, goose”, the second far in “far, far away”. Little words, numbers and code get almost nothing, and so does a word repeated a few words later. Turned up, it makes the model repeat itself sooner…","st":"tested","h":"../notes/L6_N9376/","i":7,"c":4,"m":[["duck, duck",79.0],["far, far away",58.5],["many, many",55.5],["location x3",51.5],["muy, muy",40.2],["purple, purple",19.5],["big, huge",0.95],["the the",0.02]]},{"id":"L6_N9428","k":"d","L":6,"N":9428,"t":"An esker spelling detector whose negative side reads 'skip over'","l":"This neuron fires on the closing piece of the word esker (es|ker), whatever the word means: glacial ridge, Norwegian word, Basque phrase or company name. Near-spellings like eskar or elsker get a weaker response, and glacial synonyms get none. Its negative side reads skip-family…","st":"tested","i":4,"c":7,"m":[["esker",67.5],["capitalised",66.0],["plural",59.8],["es|ke",45.8],["misspelling",25.8],["Norwegian elsker",18.0],["-ipped control",0.3],["glacial, no spelling",0.0],["überspringen",-33.0],["skip over",-41.8]]},{"id":"L6_N9814","k":"d","L":6,"N":9814,"t":"One neuron, two tenants: Severn's last letter and a polytechnic name list","l":"This neuron rents space to two unrelated tenants. Its positive side fires only on the token that finishes spelling S-E-V-E-R-N, whatever Severn names: the river, the water company, an Ontario township. Prose about the river without the word barely registers. Its negative side is…","st":"tested","i":3,"c":6,"m":[["Severn (core)",26.6],["SEVERN caps",24.9],["Sever- partial",10.2],["one-word post-92",0.6],["other river",0.4],["nonsense name",-0.2],["Brookes alone",-12.0],["Sheffield Hallam",-28.5],["Oxford Brookes",-33.0],["Bendigo",-43.2]]},{"id":"L7_N15717","k":"d","L":7,"N":15717,"t":"Using a trained model versus choosing its features: a signed axis","l":"This neuron is a signed axis over the classical machine-learning workflow. It goes positive where a fitted model is used: calling .predict, saving it with joblib or exporting PMML, or describing it predicting on test data. It goes much more negative, down to about −87, on…","st":"tested","i":5,"c":8,"m":[["save model",10.1],["predict call",10.1],["fit (halo)",2.9],["predict prose",2.6],["split (control)",0.8],["non-ML control",0.5],["tuning",-12.8],["CV splitter",-25.1],["FS triad",-41.0],["feature selection",-47.2]]},{"id":"L7_N17195","k":"d","L":7,"N":17195,"t":"Triumvirate and Lepidus pieces up, certain god names and Bode down","l":"This neuron is a word-piece detector with two unrelated poles. It fires strongly on the inner pieces of a few late-Republic words, above all 'vir' in triumvirate and 'id' in Lepidus. The spelling alone does much of the work, and Roman political context adds more, up to about…","st":"tested","i":3,"c":5,"m":[["Roman triumvirate",41.0],["bakers' triumvirate",36.2],["Lepidus",26.6],["duumvir",15.4],["Roman halo",8.8],["Crassus",5.7],["modern politics",-0.1],["control",-0.2],["Bode",-32.8],["Ares",-36.0]]},{"id":"L7_N18023","k":"d","L":7,"N":18023,"t":"A Vietnam gate that fires on the next noun, not the name","l":"L7/N18023 fires when Vietnam is in context, but not on the word Vietnam. A gate opens once Vietnam, a Vietnamese city, or Vietnamese script has appeared and stays open for several sentences. The value then lands on whatever content noun follows: university, coffee, river…","st":"tested","i":5,"c":6,"m":[["Vietnam (word)",17.6],["Ho Chi Minh City",17.6],["Hanoi (core)",17.4],["Vietnamese script",17.4],["Cambodia (halo)",3.9],["Thailand (halo)",0.7],["Colombia (control)",-0.5],["Paris (control)",-0.6],["no place",-0.6],["Korea (weak)",-0.9]]},{"id":"L7_N23572","k":"d","L":7,"N":23572,"t":"A lexical 'five elements' detector, indifferent to cosmology","l":"This neuron reads a two-word pattern, not a belief system. It fires on the word 'element(s)' when it follows five, four, fifth or classical, in any language and any subject, so 'the five elements of hip hop' scores the same as the Wu Xing. Element nouns in a list get a smaller…","st":"tested","i":4,"c":8,"m":[["five classical",60.5],["five",42.8],["hip-hop five",27.2],["four",23.2],["array five",22.6],["seven",12.2],["five chemical",9.9],["six",9.8],["no numeral",8.9],["key (rhetorical)",7.4]]},{"id":"L7_N23674","k":"d","L":7,"N":23674,"t":"Creole language gate: connectives fire positive, pronoun 'li' fires negative","l":"This neuron multiplies a language gate by a word-specific sign. The gate stays shut on English and opens once text reads as Haitian or Mauritian Creole, Malagasy or romanized Hokkien. Inside that gate most tokens go positive, with connective pieces like e|pi and no|ho the…","st":"tested","i":5,"c":5,"m":[["French",32.2],["nonsense",31.9],["Malagasy",25.5],["Jakarta",19.9],["Yoruba (halo)",19.1],["no cue",-26.4],["English",-30.5],["Mauritius",-30.8],["Haiti",-34.0],["Kreyòl cue",-37.0]]},{"id":"L7_N9251","k":"d","L":7,"N":9251,"t":"Long shots, shoo-ins and underdogs: the words for who’s expected to win","l":"An early neuron that fires at the end of the stock phrases for a contestant’s chances: long shot, underdog, odds-on, front-runner, shoo-in, dark horse. Good odds or bad. It’s the words, not the idea: “not expected to win” barely moves it. It tells a race favorite from a favorite…","st":"tested","h":"../notes/L7_N9251/","i":4,"c":5,"m":[["a long shot",44.0],["big underdogs",36.2],["odds-on",32.0],["camera long shot",24.1],["race favourites",23.4],["a shoo-in",21.6],["not expected",4.22],["favorite color",0.01]]},{"id":"L8_N1000","k":"d","L":8,"N":1000,"t":"Completes famous education-agency names, negative for other names in schooling text","l":"This layer-8 neuron fires on the token that completes a recognized education-sector name. It peaks on the last piece of famous agency acronyms such as NCES, MEC, YÖK and NEP, and fires more weakly on compounds like 'education policy'. Education context opens the gate. The up…","st":"tested","i":4,"c":7,"m":[["NCES expanded",21.2],["bare NCES",19.4],["MEC expanded",19.4],["lesser acronym",8.5],["edu compound",8.1],["'education' word",3.0],["non-edu compound",0.6],["nonsense control",-0.1],["non-edu acronym",-1.4],["edu surname",-5.3]]},{"id":"L8_N24564","k":"d","L":8,"N":24564,"t":"A cross-lingual date-fruit detector that dips on bell peppers","l":"This neuron recognizes the date fruit. It fires on the word for dates in English, French, Italian, Malay, Arabic script and Latin botany, most strongly when a dried-fruit list or iftar context backs the fruit sense (up to about +50). Prunes and figs get a weaker echo, and…","st":"tested","i":5,"c":9,"m":[["French dattes",50.0],["dates",47.5],["Medjool dates",40.5],["prunes (halo)",24.2],["Malay kurma",23.5],["figs (halo)",17.2],["calendar dates",12.9],["olives",3.9],["carrots",0.7],["bell pepper",-10.7]]},{"id":"L8_N4083","k":"d","L":8,"N":4083,"t":"A neuron for how far a produced sound carries","l":"This unit fires on the word that names how far a produced sound reaches, 'projection', 'project' or 'carry', once earlier text has named a voice, instrument or speaker. The same word for a film, map or budget stays near zero. Related quality nouns form a ladder, with…","st":"tested","i":5,"c":7,"m":[["singing voice",16.1],["brass",13.1],["cello",12.4],["stage voice",9.1],["nonsense",6.0],["phone speaker",5.6],["bullfrog",4.9],["film (control)",2.8],["budget (control)",1.3],["map (control)",-0.3]]},{"id":"L8_N5247","k":"d","L":8,"N":5247,"t":"One gate, two words: 'action' in use versus the minor coin","l":"This neuron fires on the mass noun 'action' when it means being in play: a quarterback who saw action, a ship out of action, a bill in action. A verb of receiving such as 'saw' roughly doubles it, and other spells of engagement like screen time or sea duty fire weaker. Its…","st":"tested","i":5,"c":7,"m":[["action",16.9],["sea duty",11.2],["playing time",10.7],["use",10.6],["plural actions",5.9],["combat",4.9],["attention",3.3],["nonsense",0.9],["cents",-7.2],["paise",-12.1]]},{"id":"L9_N15817","k":"d","L":9,"N":15817,"t":"A name detector for Koran and Mahomet, with a faint scholarly echo","l":"This neuron is mainly a negative detector for a learned set of religious names. It responds most strongly to the Koran in any spelling and to the old European name Mahomet, reaching about -20. Weaker members are Mormon, Christian Science and the Ahmadiyya, and these fire even…","st":"tested","i":4,"c":6,"m":[["Ramsay (+)",0.9],["control",0.2],["Korea (lookalike)",0.1],["Bible",-0.4],["Sikh scripture",-7.2],["Mormon",-8.5],["Alcoran",-9.8],["Chr. Science",-10.9],["Mahomet",-11.6],["Koran",-17.0]]},{"id":"L9_N6153","k":"d","L":9,"N":6153,"t":"A probability-only axis: event relations versus sample-space outcomes","l":"This neuron only works inside probability text, where it acts as a signed axis. Words that relate events push it positive: dependent, independent, the bar in P(A|B), named rules. The sample space and its points push it much further negative: ω, Ω, 'outcomes', 'sample points'. A…","st":"tested","i":5,"c":8,"m":[["dependent",4.3],["cond. bar",3.4],["variance (halo)",2.3],["named rule",2.0],["expected value",0.8],["no-concept control",-0.2],["results",-3.8],["sample space",-8.2],["elementary events",-10.2],["outcomes",-12.3]]},{"id":"L0_E100_N436","k":"m","L":0,"E":100,"N":436,"t":"A classical-language expert's neuron that mostly reads the character 設","l":"This neuron sits in an expert that the router reserves for classical-language text: classical Chinese, Latin and Homeric Greek. Within that text it acts like a small lookup table of specific tokens. The character 設 ('establish') is the dominant entry, at about +1.2. A few tokens…","st":"tested","i":3,"c":2,"m":[["設 core",1.2],["设 simplified",1.2],["secondary entry",1.0],["secondary entry",0.5],["build, weak",0.4],["synonym control",-0.0],["synonym control",-0.0],["plain control",-0.0],["negative entry",-0.3],["negative entry",-0.5]]},{"id":"L0_E11_N16","k":"m","L":0,"E":11,"N":16,"t":"A dead neuron: zero weights, zero output, labels from ties","l":"This neuron is dead. All of its weights sit at the smallest representable value, so its input and output vectors are effectively zero. It outputs exactly zero on millions of text and image tokens, and pushing it to +500 changes no prediction. The 'math connective' labels…","st":"tested","i":1,"c":8,"m":[["math opener",-0.0],["math opener",-0.0],["opener variant",-0.0],["opener variant",-0.0],["connective",0.0],["prose-like",-0.0],["plain prose",-0.0],["control: news",-0.0],["control: data",-0.0]]},{"id":"L0_E161_N309","k":"m","L":0,"E":161,"N":309,"t":"A router-gated detector for the Chinese word 保养 (upkeep)","l":"This layer-0 neuron is a token detector for the Chinese word 保养 ('upkeep, servicing'). It fires strongly on that word in every domain, gives a weaker response to drill words (演习, ' drills'), and ignores true synonyms such as 维护. So it tracks spelling, not meaning. Whether it…","st":"tested","i":3,"c":8,"m":[["保养 (core)",4.0],["traditional form",3.5],["drill",2.3],["halo synonym",1.7],["drill variant",1.5],["training",0.8],["control: maintain",0.0],["control: health care",-0.0],["control: clean",0.0],["negative pole",-0.1]]},{"id":"L0_E188_N267","k":"m","L":0,"E":188,"N":267,"t":"A layer-0 neuron that fires negative on the word 'select'","l":"This early-layer neuron is a detector for one word: 'select'. When the bare token 'select' appears, in any casing, as a button label, in SQL or in prose, it swings strongly negative, reaching about -3. Forms like 'selected' respond at roughly half strength, and a repeated…","st":"tested","i":3,"c":9,"m":[["derivation",-0.0],["halo (+)",0.0],["synonym",-0.0],["lookalike",-0.0],["control",0.0],["inflection",-1.2],["core",-2.2],["capitalized",-2.5],["all caps",-3.6]]},{"id":"L0_E190_N361","k":"m","L":0,"E":190,"N":361,"t":"A dead neuron whose examples only trace its expert's router","l":"This neuron is dead. Its input and output weights are about 1e-38, so it outputs a signed zero on every token, text or image. The labels about 'digits inside numbers' and 'LaTeX' come from ties in the ranking, not from the neuron. What the examples really show is where its…","st":"tested","i":1,"c":1,"m":[["2nd digit in group",0.0],["group digits",0.0],["1st digit after comma",0.0],["LaTeX command",0.0],["problem opener",0.0],["geometry prose",0.0],["code (control)",0.0],["prose (control)",0.0],["Chinese (control)",0.0]]},{"id":"L0_E247_N368","k":"m","L":0,"E":247,"N":368,"t":"A spelling detector: glued 'st' pieces up, 'men' and 'ment' down","l":"This early neuron reads spelling. It fires when a short lowercase piece beginning with 'st' is glued, with no space, onto a preceding word, as in the German compound Generalstab or the run-together recommendationsto. It swings the other way on glued 'men' (selectmen) and 'ment'…","st":"tested","i":3,"c":8,"m":[["glued stab",2.1],["glued sth",1.8],["glued str",1.3],["est halo",1.0],["s+other",0.1],["spaced stab",0.0],["control suffix",0.0],["control word",-0.0],["glued ment",-0.9],["glued men",-1.7]]},{"id":"L0_E45_N445","k":"m","L":0,"E":45,"N":445,"t":"则 up, 通 down, but only right after a personal name","l":"This layer-0 neuron reads one token's identity. The token 则 pushes it strongly positive and 通 pushes it about equally negative. Almost everything else leaves it near zero. Its expert only receives 则 or 通 when the previous token is part of a personal name. So in real text it…","st":"tested","i":3,"c":10,"m":[["则 (core)",2.4],["merged 则在",0.7],["却 connective",0.3],["而 connective",0.2],["说 (control)",0.1],["也 (control)",0.0],["的 (control)",0.0],["merged 通则",-0.2],["通 (negative)",-2.1]]},{"id":"L0_E88_N183","k":"m","L":0,"E":88,"N":183,"t":"A dead neuron borrowing its expert router's taste for multiplication","l":"This neuron is dead. Its input and output weights in expert 88 sit at the numerical floor, so it outputs zero on every routed text and image token, and forcing it to a large value changes nothing downstream. Its 'multiplication' label comes from elsewhere: expert 88's router…","st":"tested","i":1,"c":1,"m":[["star",-0.0],["power",-0.0],["\\times",-0.0],["word times",-0.0],["multiplied by",-0.0],["plus (control)",-0.0],["divide (control)",-0.0],["prose (control)",-0.0]]},{"id":"L0_E92_N246","k":"m","L":0,"E":92,"N":246,"t":"A word weigher: ' equ' up, 'store' down","l":"This layer-0 neuron gives a signed weight to a few specific words when they sit in a verb slot. The word piece ' equ' (as in equidistant or equate) pushes it up. The verb 'store' pushes it down. Synonyms such as save, keep and record do almost nothing, so the neuron tracks the…","st":"tested","i":3,"c":8,"m":[["equ(ate)",1.1],["complete",0.5],["develop",0.5],["answer",0.3],["save (control)",-0.0],["find (control)",-0.1],["hold",-0.1],["create (control)",-0.3],["represent",-0.3],["store",-0.8]]},{"id":"L0_E99_N105","k":"m","L":0,"E":99,"N":105,"t":"A dead neuron: all-zero weights, zero output everywhere","l":"This neuron does not detect anything. Its input and output weights are all exactly zero, so it outputs zero on every token, text or image, and writes nothing to the model. Its apparent 'math prompt' examples come from arbitrarily ranking millions of ties at zero. It is one of…","st":"tested","i":1,"c":5,"m":[["Find",-0.0],["Compute",-0.0],["Determine",-0.0],["Calculate",0.0],["Give",-0.0],["Explain",0.0],["question",-0.0],["framing",-0.0],["control",0.0],["control",0.0]]},{"id":"L10_E11_N415","k":"m","L":10,"E":11,"N":415,"t":"Free and independent versus owned and operated: a signed phrase-completion neuron","l":"This neuron watches the word that completes a set phrase, such as the second half of 'free and ___'. That slot always opens the gate, and the word chosen sets the sign. Status and criterion words in sovereignty, doctrine or rule sentences ('a free and independent state') push it…","st":"tested","i":4,"c":6,"m":[["close variant",0.9],["core status",0.8],["status variant",0.0],["off-script halo",-0.0],["managing verb",-0.0],["managing verb",-0.0],["managing verb",-0.0],["control",-0.0],["control",0.0],["control",0.0]]},{"id":"L10_E125_N303","k":"m","L":10,"E":125,"N":303,"t":"Newlines written by code versus the 'Here's your code' break","l":"This neuron is a signed dial on line breaks. It turns positive where code is about to write a newline: most clearly the typed '\\n' at the head of a returned string, and also the break that opens a function body. It turns negative on the blank line after a chatty handover like…","st":"tested","i":5,"c":8,"m":[["def header",1.9],["indented def",1.2],["class header",0.3],["loop header",-0.0],["third person",-0.0],["control label",0.0],["impersonal",-0.2],["eager, no code",-0.8],["no 'Of course'",-1.0],["eager + code",-2.0]]},{"id":"L10_E133_N117","k":"m","L":10,"E":133,"N":117,"t":"A neuron that fires on 'Scrabble value' and predicts 'of'","l":"This neuron is a narrow word detector. It fires strongly on the noun 'value' (or 'score') when it closes the phrase 'Scrabble value' on word-dictionary pages. There it helps predict 'of' in the 'Scrabble value of WORD' line, and removing it roughly halves that prediction.…","st":"tested","i":3,"c":8,"m":[["core",6.1],["word between",5.3],["plural",4.7],["crime halo",4.4],["possessive",3.7],["other game",2.8],["score",2.4],["other board game",0.7],["negative",0.1],["no game",0.0]]},{"id":"L10_E135_N390","k":"m","L":10,"E":135,"N":390,"t":"Code reference completion versus type names: a signed syntax unit","l":"This layer-10 neuron is a signed code-syntax detector. It fires on the token that completes a value-side reference in code: the ')' closing a call over names, or the next segment of a dotted or namespaced path. It turns negative on user-defined type names in type slots…","st":"tested","i":5,"c":8,"m":[["call close",1.2],["dotted path",0.6],["mid-call comma",0.3],["control: number",0.2],["call close 2",0.0],["module path",-0.0],["control: comment",0.0],["import alias",-1.1],["exception name",-1.4]]},{"id":"L10_E149_N314","k":"m","L":10,"E":149,"N":314,"t":"A neuron for the first path join in Asymptote drawings","l":"This neuron in Qwen3.6-35B-A3B recognizes the first line segment of a figure drawn in Asymptote. It fires on the '--' that joins two points inside draw(), and it barely responds to the same symbol in an assignment or an ordinary function call. Each further join weakens it, to…","st":"tested","i":5,"c":6,"m":[["filldraw join",3.4],["draw join",3.0],["third join",3.0],["tuple join",2.2],["path assign",1.9],["plain assign",0.8],["other call",0.5],["comment",0.5],["C decrement",0.0]]},{"id":"L10_E155_N335","k":"m","L":10,"E":155,"N":335,"t":"A neuron that marks renaming phrases: opened negative, completed positive","l":"This neuron tracks a phrase that renames something just mentioned, as in 'Dunhuang, a city' or 'this pope'. It fires positive on the noun that completes the renaming, especially in apposition and 'X called Y' constructions. It fires negative on the classifier or modifier that…","st":"tested","i":6,"c":6,"m":[["demonstrative head",0.2],["connective control",-0.0],["bare head",-0.4],["pronoun control",-0.9],["apposition head",-1.9],["head variant",-1.9],["classifier opener",-1.9],["modifier onset",-1.9],["关于 onset",-2.5]]},{"id":"L10_E180_N225","k":"m","L":10,"E":180,"N":225,"t":"A neuron for removing zero, opposed by independence and disjointness","l":"This neuron marks the moment a math text says an element, usually zero or the identity, has been taken out of a set: 'the zero vector removed'. Swapping in 'included' or 'added' turns it off. Its larger negative pole comes from a few relations that keep structures apart…","st":"tested","i":5,"c":4,"m":[["zero removed",1.0],["zero excluded",0.9],["ordinary removed",0.3],["included",0.3],["added",-0.1],["control",-0.1],["lower bound",-0.2],["separator",-0.8],["linearly disjoint",-2.1],["cond. independence",-5.2]]},{"id":"L10_E195_N300","k":"m","L":10,"E":195,"N":300,"t":"A signed word list: Hārūn and mow up, Napoleonic Wars down","l":"This neuron works like a short signed word list, not a detector of any general concept. A few specific entries push it up: the Arabic name Hārūn, the verb mow, rare literary names, and '+' or arrows inside repeated templates. A separate set pushes it down: 'Napoleonic Wars', 'is…","st":"tested","i":2,"c":2,"m":[["Hārūn",1.0],["full name",1.0],["no macrons",0.4],["Hrūn",-0.0],["Q control",-0.0],["M control",-0.0],["B control",0.0],["other macron name",0.0],["Rūmī",-0.0]]},{"id":"L10_E229_N346","k":"m","L":10,"E":229,"N":346,"t":"A neuron for the B in Royal Society B","l":"This neuron fires on the letter that names one series of a lettered journal: the B in 'Philosophical Transactions of the Royal Society B' is its strongest case. It responds to titles the model has memorized rather than to any letter after any journal, so Physical Review B stays…","st":"tested","i":4,"c":6,"m":[["Royal Soc. B",3.0],["Mater. Chem. B",3.0],["J. Vac. B",2.5],["Chromatogr. B",2.3],["J. Phys. B",1.1],["invented title",0.9],["Phys. Rev. B",-0.0],["no letter",0.0],["room label",0.0],["acronym",-1.8]]},{"id":"L10_E231_N99","k":"m","L":10,"E":231,"N":99,"t":"A neuron for 'you're given two sentences' in task prompts","l":"This neuron fires on the noun that completes a task instruction handing the reader text or symbols: the 'sentences' in 'In this task, you're given two sentences'. The router supplies the frame and the neuron supplies the noun preference. A separate route fires on a few memorized…","st":"tested","i":4,"c":5,"m":[["sentences (core)",0.9],["statements",0.9],["strings",0.8],["integers",0.7],["words (weak)",0.7],["numbers",0.6],["cars (control)",0.1],["Hazaras (neg)",0.0],["apples (control)",0.0],["arrays (neg)",-0.3]]},{"id":"L10_E56_N128","k":"m","L":10,"E":56,"N":128,"t":"A signed key reader: team and capital versus instance-of","l":"This neuron only works on keys and labels in structured text, such as tables, JSON and infoboxes; in ordinary prose it stays silent. There it gives the key word a sign. Civic and sports field names like team, coach, capital and stadium read positive, and the memorized…","st":"tested","i":4,"c":6,"m":[["coach team",1.0],["team",0.9],["stadium",0.4],["school",0.4],["capital",0.0],["control: the",-0.0],["control: prose",-0.1],["social relation",-0.3],["instance",-1.0],["hasPart",-1.3]]},{"id":"L10_E77_N206","k":"m","L":10,"E":77,"N":206,"t":"A neuron that recognizes school initialisms, but only in school contexts","l":"This neuron fires on a short initialism, such as TMS, only when two conditions hold: the text treats it as a school that people attend, and the letters look like a school name. Either condition alone gives little. 'The students at TMS' reads about 9, while 'engineers at TMS' and…","st":"tested","i":6,"c":7,"m":[["intl school",5.9],["school (core)",5.7],["high school",3.5],["elementary",2.6],["famous agency",1.6],["odd letters",1.4],["concept acronym",1.3],["famous company",1.1],["nonsense letters",0.6],["plain word",0.0]]},{"id":"L10_E77_N314","k":"m","L":10,"E":77,"N":314,"t":"A neuron that reads SOLID as design principles, against JAGS","l":"This neuron is close to a one-word detector: it fires on the piece that completes uppercase SOLID when the word names the object-oriented design principles, in English, Spanish or French, with or without code around it. Other senses of SOLID, lowercase 'solid' and sibling…","st":"tested","i":4,"c":8,"m":[["dotted SOLID",0.7],["SOLID (core)",0.6],["halo name",0.5],["halo name",0.2],["sibling acronym",0.1],["JAGS (negative)",0.1],["Titlecase",-0.0],["lowercase",0.0],["sibling acronym",-0.1],["non-word acronym",-0.5]]},{"id":"L10_E8_N174","k":"m","L":10,"E":8,"N":174,"t":"A neuron for 'team approach', set against workflow-and-optimize vocabulary","l":"This neuron tracks the idea of the team as a professional way of working. It fires on 'team' in care or professional prose and fires hardest on the method word that follows, as in 'a team approach'. Its opposite is a narrow set of process words: 'workflow' and the optimize…","st":"tested","i":4,"c":6,"m":[["care team",3.1],["team approach",3.0],["committee approach",2.4],["team effort",2.1],["integrated approach",0.6],["control: checkup",0.6],["soccer team",-0.0],["control: diet",-0.0],["optimized",-1.2],["workflow",-2.2]]},{"id":"L11_E111_N406","k":"m","L":11,"E":111,"N":406,"t":"How X is made versus X to try: a signed neuron","l":"This neuron reads what a named thing is being asked about. It is positive on the word naming how the thing comes about or works, such as 'how cheese is made', 'the election process' or 'chord construction'. It is negative when the thing is framed as an option to sample, as in…","st":"tested","i":5,"c":6,"m":[["declarative (control)",2.1],["how made",1.9],["how produced",1.3],["process noun",0.8],["how works",0.4],["wh-question",0.4],["history (control)",0.2],["definition",0.1],["to order",-1.1],["to try",-1.8]]},{"id":"L11_E129_N356","k":"m","L":11,"E":129,"N":356,"t":"A signed lookup of memorized citation strings, with no measurable effect","l":"This neuron is a small signed lookup table of memorized citation strings. It fires positive on an inner piece of 'newadvent.org/cathen' and 'Biochem Biophys Res Commun', and fires negative on the digit after 'doi:10.1007/s00' and on ' TES' in 'TESL Journal'. One shared gate…","st":"tested","i":2,"c":2,"m":[["BBRC (core)",0.8],["cathen URL",0.6],["BBRC partial",-0.0],["near variant",-0.0],["JBC control",-0.0],["Lancet control",-0.0],["Cell Metab control",-0.0],["TESL (neg)",0.0],["Springer DOI (neg)",-0.6]]},{"id":"L11_E158_N396","k":"m","L":11,"E":158,"N":396,"t":"A neuron that weighs 'some context' against 'seven drugs'","l":"This neuron tracks the grammatical number of the noun phrase being built. It fires positively on the noun right after 'some' when that noun is singular or mass, as in 'For some context,'. It turns negative inside plural phrases counted by a number or 'other', as in 'seven…","st":"tested","i":4,"c":6,"m":[["some + singular",1.4],["framing noun",1.2],["framing noun",0.6],["partitive",0.0],["bare noun",0.0],["some + plural",-0.1],["counted plural",-0.2],["other + plural",-0.2],["numeral plural",-0.6]]},{"id":"L11_E210_N198","k":"m","L":11,"E":210,"N":198,"t":"Negative on 'they looked' after people made or handled things","l":"This neuron fires negatively on the verb right after 'they' when a group of people has just made, bought or handled objects: 'they looked', 'they came out', 'they turned out'. The handling verb sets the strength, so 'made pillowcases' goes deepest (about -7.6) and 'ran a…","st":"tested","i":5,"c":6,"m":[["enjoyed (control)",-0.1],["smiled (control)",-0.3],["ran (control)",-0.6],["tie-dyed",-3.1],["saw (halo)",-3.2],["bought",-4.4],["graded exams",-6.0],["washed",-6.1],["made",-6.4]]},{"id":"L11_E33_N56","k":"m","L":11,"E":33,"N":56,"t":"A neuron that recognizes laptop sound and network chip vendors","l":"This neuron recognizes PC chipmaker names in their hardware sense. It fires on the inner pieces of audio-codec and network-chip vendors such as SigmaTel, IDT, Realtek, Conexant and Broadcom, and on the first letters of their part numbers. The same letters in another sense stay…","st":"tested","i":6,"c":6,"m":[["SigmaTel",9.4],["IDT",7.9],["Realtek",2.7],["Conexant",2.6],["Broadcom",2.5],["Atheros (weak)",1.4],["brand halo",0.9],["Intel control",0.6],["invented vendor",0.1],["CLI flag",-2.2]]},{"id":"L11_E79_N293","k":"m","L":11,"E":79,"N":293,"t":"What is ruled out: excluded items versus outsiders' access","l":"Inside an expert that handles stated or sought absences, this neuron sorts what is being excluded. When a rule, design or effort rules out a structure or act, such as a graph that must not contain a cycle, a glitch-free interface or trying not to cry, it fires positive. When the…","st":"tested","i":5,"c":7,"m":[["excluded item",0.5],["-free compound",0.4],["no recording",0.4],["no-X constraint",0.2],["control: affirmative",0.0],["control: neutral",-0.0],["ban halo",-0.3],["no-logs",-0.3],["susceptibility",-1.1],["unappealing",-1.8]]},{"id":"L11_E79_N399","k":"m","L":11,"E":79,"N":399,"t":"A neuron that tells 'none of them' from 'leave out any'","l":"This neuron reads the word 'any' (and relatives like 'none', 'either', 'aucun') and signals which sense is meant. It goes positive when the sentence says no member of a set qualifies, as in 'has not paid any of the rent'. It goes negative when 'any' is the object of a verb of…","st":"tested","i":7,"c":6,"m":[["not paid",7.5],["never",6.6],["hardly",5.4],["denied",2.8],["not omit",0.1],["control: can",0.0],["control: read",0.0],["new to",-1.8],["overlooked",-2.3],["omit",-5.0]]},{"id":"L11_E96_N63","k":"m","L":11,"E":96,"N":63,"t":"A signed noun reader: field and polity versus measured quantity","l":"This neuron reads the head noun of a noun phrase in expository prose, and it is signed. It goes positive when the passage expects a field of study, a polity in a union, or a pulling force. It goes negative when the passage expects a measured quantity ('correlation measures the…","st":"tested","i":3,"c":2,"m":[["Kingdom (core)",0.0],["Republic",0.0],["Crown",0.0],["countries",0.0],["strength",0.0],["speed (neg)",-0.0],["women (neg)",-0.0],["two (control)",0.0],["same (control)",0.0],["parliaments (halo)",-0.1]]},{"id":"L12_E117_N492","k":"m","L":12,"E":117,"N":492,"t":"Mathematics versus Internet Policy: a neuron for scholarly name endings","l":"This neuron watches the last word of a scholarly name. When that word is 'mathematics' naming a field, especially closing a list of sciences or an institution name, it fires positively. When the word is 'Policy' or 'Politics' right after 'Internet' in a journal title like…","st":"tested","i":3,"c":5,"m":[["applied math",0.0],["mathematics",0.0],["physics",0.0],["chemistry",-0.0],["control",0.0],["public policy",0.0],["web policy",-0.1],["internet ethics",-1.1],["internet policy",-2.9],["internet politics",-3.1]]},{"id":"L12_E15_N488","k":"m","L":12,"E":15,"N":488,"t":"A neuron for premises reported as 'given' to the solver","l":"This neuron marks the moment a math solution reports a fact as handed to the solver. It fires hardest on the word that completes 'is given as' or 'given to be', just before the value arrives, and on 'Given:' headers that open a list of premises. Other reporting verbs ('We are…","st":"tested","i":5,"c":8,"m":[["given as",4.5],["given to be",3.0],["specified as",1.5],["given:",1.4],["general law",1.4],["stated as",1.3],["known to be",1.1],["listed as",0.6],["assumed",0.2],["plain is",0.1]]},{"id":"L12_E165_N124","k":"m","L":12,"E":165,"N":124,"t":"Post and telephone services up, their phone numbers down","l":"This neuron tracks the post and telephone service as an institution. It fires on words like postal, post office, telegraph, the telephone exchange and the phone book, in many languages, and peaks in encyclopedic prose about national post-and-telephone services. It turns negative…","st":"tested","i":5,"c":6,"m":[["savings bank",1.0],["post office",0.9],["telegraph",0.8],["phone book",0.6],["control: railway",0.4],["phone exchange",0.2],["DNS",0.1],["postal",0.0],["control: parks",0.0],["phone digits",-1.1]]},{"id":"L12_E187_N6","k":"m","L":12,"E":187,"N":6,"t":"A neuron that likes cited titles and dislikes 'and how can'","l":"This neuron goes up on a small set of nouns, such as Empire, World and armour, when they close a phrase in reference-style writing, and it peaks on the last word of a cited title. It goes most negative at the 'can' in follow-up questions like '…, and how can they be fixed?'.…","st":"tested","i":3,"c":5,"m":[["Empire (core)",0.0],["World",0.0],["Range",0.0],["City (halo)",0.0],["Meaning (halo)",0.0],["Republic (control)",0.0],["Revolution (control)",0.0],["Brotherhood (control)",0.0],["and how can",-0.2]]},{"id":"L12_E233_N178","k":"m","L":12,"E":233,"N":178,"t":"Knocked off its rest: bails, bars, and toppled vases","l":"This neuron responds to a standing or perched object that has just been struck, and it fires as soon as the object is named after the contact verb. Two signals add: the physical knock itself, and a sports-obstacle setting such as cricket bails or a jump bar. Together they…","st":"tested","i":6,"c":8,"m":[["cricket bails",17.1],["block tower",11.4],["jump bar",8.7],["vase",5.2],["ball away",4.2],["control: score",3.9],["show-jump rail",2.1],["bunker",0.8],["control: price",0.0]]},{"id":"L12_E237_N49","k":"m","L":12,"E":237,"N":49,"t":"A neuron for 'you learn it by doing it'","l":"This neuron marks the moment a text names hands-on doing (practice, work, labour, continual effort) as the way a skill is acquired. It fires most in Chinese advice such as 这些技巧都需要在实践中不断学习. Swapping in study, observation or theory largely switches it off. A weak opposite pole…","st":"tested","i":5,"c":8,"m":[["labour",6.7],["practice (core)",5.4],["work",5.2],["continual practice",5.0],["use (halo)",4.2],["theory (control)",3.3],["real life (neg)",2.6],["research (halo)",1.4],["wedding (control)",0.0]]},{"id":"L12_E249_N460","k":"m","L":12,"E":249,"N":460,"t":"A neuron that spots the first number of a new block","l":"This neuron watches lists of numbers that share leading digits. When the leading digits change and a new block begins (id21 after id11, id12, id13), it fires on the new number's last digit, most strongly when that digit repeats where the old block started. The same pattern…","st":"tested","i":5,"c":6,"m":[["restart, skip",4.8],["restart echo",3.8],["partial echo",2.7],["skip in block",1.9],["big jump",1.3],["figure locator",0.3],["page locator",-0.0],["next in run",0.0],["no number",-0.0],["repeat",-0.1]]},{"id":"L12_E24_N277","k":"m","L":12,"E":24,"N":277,"t":"A neuron that fires on memorized word-in-frame pairs","l":"This neuron has no topic. It responds to particular word-in-frame pairs it seems to have memorized. 'adding' after 'Your task is to complete the implementation by' fires strongly, and so do 'play' closing 'live, learn, work' and '/div' in XPath strings. 'problem' after 'classic…","st":"tested","i":3,"c":2,"m":[["finish impl.",0.5],["implementation",0.4],["code",0.4],["class",0.3],["extend",0.3],["recipe",0.3],["sculpture",0.1],["soup",0.0],["room",-0.0]]},{"id":"L12_E32_N278","k":"m","L":12,"E":32,"N":278,"t":"A neuron for things that ended, not things absent","l":"In Qwen3.6-35B-A3B, one neuron at layer 12 responds to things that ended or came to nothing. It is strongest on the word that closes a story of a project that fizzled, a request turned down or a club wound up, and it grows when such failures accumulate. It turns negative when…","st":"tested","i":7,"c":6,"m":[["stopped",1.7],["gave up",1.7],["abandoned",1.3],["on hold",1.2],["halo: barely",0.9],["never finished",0.5],["avoidance",0.4],["control: finished",0.0],["control: published",-0.0],["control: revised",-0.0]]},{"id":"L12_E34_N350","k":"m","L":12,"E":34,"N":350,"t":"A neuron that marks the Bronze Age era and 'Athens Institute'","l":"This neuron is mainly a negative detector. It fires most strongly when 'Bronze' names the archaeological era, as in 'the Late Bronze Age', in English, French, German or Greek. It stays quiet for bronze as a metal. A second, memorized trigger is the name 'Athens Institute'. Other…","st":"tested","i":4,"c":5,"m":[["Roman (control)",0.1],["decade (control)",0.1],["summer (control)",0.0],["Heroic",-0.2],["Late Helladic",-1.1],["Iron",-1.1],["2nd millennium",-1.3],["Chalcolithic",-1.4],["Late Bronze",-3.3],["Early Bronze",-3.8]]},{"id":"L12_E48_N185","k":"m","L":12,"E":48,"N":185,"t":"A neuron for laws that deputize anyone to sue","l":"This neuron in layer 12 of Qwen3.6-35B-A3B is essentially a negative detector for laws that let any member of the public sue and collect money. Texas's SB8 bounty clause is the archetype and drives it to about -19. Texas lawmaking, Prop 65 bounty plaintiffs and old…","st":"tested","i":7,"c":7,"m":[["control: company",-2.8],["the state",-3.7],["injured party",-3.8],["control: dog",-4.0],["anyone",-6.6],["any resident",-8.2],["abortion halo",-8.6],["bounty hunters",-9.1],["informer",-9.3],["private citizen",-11.2]]},{"id":"L12_E68_N126","k":"m","L":12,"E":68,"N":126,"t":"A neuron that weighs the user's choice against the posed problem","l":"This neuron weighs one kind of object against another. It fires on the noun in an input-request slot, and its sign says what that object is. It goes positive when the noun names something the user will choose ('enter your choice', about +2). It goes negative when the noun names…","st":"tested","i":4,"c":7,"m":[["choice",1.5],["pick",1.3],["selection",0.9],["option",0.6],["preference",0.0],["number (control)",0.0],["name (control)",0.0],["question",-0.0],["problem",-0.2]]},{"id":"L12_E79_N156","k":"m","L":12,"E":79,"N":156,"t":"A neuron for 'find out more', minus memorized string pieces","l":"This is a weak, mostly lexical neuron. In prose it fires on a bare 'more' that closes a clause about finding out, learning or buying ('to find out more'), and more weakly on other clause-final nouns in explanatory text. In code it also fires on the Tailwind margin class and on…","st":"tested","i":2,"c":2,"m":[["find out more",0.2],["even more",0.2],["learn more",0.1],["discover more",0.1],["buy more",-0.0],["information",-0.0],["earn more (ctrl)",-0.0],["less (ctrl)",0.0],["the answer (ctrl)",0.0],["/latest (neg)",0.0]]},{"id":"L12_E7_N148","k":"m","L":12,"E":7,"N":148,"t":"A neuron for deliberate effort framed by 'once', 'if' or 'who'","l":"This neuron fires when a sentence frames someone's deliberate effort or search inside a 'once', 'when', 'if' or 'who' clause. Examples are 'once you start looking for' and 'even those who actively try'. The peak falls on the effort word or on the 'to'/'for' that opens the goal.…","st":"tested","i":5,"c":6,"m":[["deliberately set out",8.4],["looking for",7.6],["actively try",7.6],["looking at",7.6],["concerted effort",4.5],["set out (plain)",3.4],["plain action",0.8],["errand",0.0],["degree phrase",-0.2]]},{"id":"L13_E100_N310","k":"m","L":13,"E":100,"N":310,"t":"A neuron for the defining attribute in 'things that have X'","l":"This neuron fires on the noun that completes a class-defining attribute, when text picks out things by what they have, use or are based on: 'browsers that use the Blink [engine]' or 'cities with a large Jewish [community]'. The same words stated as a fact about one thing stay…","st":"tested","i":4,"c":6,"m":[["browser engine",7.5],["engine variant",5.7],["city community",3.2],["app data (halo)",3.2],["car engine",1.6],["laptop CPU (halo)",0.8],["open set",0.0],["plain fact",0.0],["no frame",0.0],["hedge",-1.0]]},{"id":"L13_E119_N244","k":"m","L":13,"E":119,"N":244,"t":"Contracts and catchphrases positive, math problem statements negative","l":"This neuron tracks register more than any single word. Contract boilerplate keeps it weakly positive, and the closing word of a few memorized phrases ('got time for that', lorem ipsum) sets off sharp positive peaks. Math and programming problem statements push it negative, and…","st":"tested","i":3,"c":5,"m":[["body",0.7],["human body",0.5],["organism",0.2],["tissue",0.0],["plant",0.0],["grid (spec)",0.0],["matrix (spec)",-0.0],["table",0.0],["control",-0.0],["control",0.0]]},{"id":"L13_E124_N244","k":"m","L":13,"E":124,"N":244,"t":"Counted back from today versus named as a notable date","l":"This neuron tracks how a story places an event in time. It fires strongly when a past event is counted back from the present, as in 'Ten years ago today, the tanks crossed into Iraq', and its peak lands at the end of the event clause. It goes negative when the same kind of day…","st":"tested","i":6,"c":7,"m":[["ago today",9.3],["exactly ago",8.1],["bare ago",5.9],["relative span",4.0],["almanac",3.8],["plain date",1.1],["control",0.4],["control",0.3],["the day that",-1.5],["notable date",-2.7]]},{"id":"L13_E150_N63","k":"m","L":13,"E":150,"N":63,"t":"Debt clauses push it up; the word 'financial' pushes it down","l":"Inside money text, this neuron splits by sign. It goes positive across the clause that says what a named money stock is made of, where it went or who supplied it. The money noun sets the strength: debt and spending sit above revenue, and a figure is not needed. It goes negative…","st":"tested","i":4,"c":6,"m":[["debt",2.1],["budget",1.5],["aid request",1.1],["spending",0.5],["electricity (halo)",0.4],["population (ctrl)",0.1],["revenue",0.0],["land (halo)",0.0],["coastline (ctrl)",0.0],["'financial' (neg)",-1.7]]},{"id":"L13_E171_N492","k":"m","L":13,"E":171,"N":492,"t":"A memorized-slot neuron that lights up on ArgumentParser(description","l":"This neuron behaves like a small lookup table of memorized code slots, not a general rule. It fires most strongly on the keyword that opens argparse's ArgumentParser call, usually 'description'. It responds more weakly to a few other familiar keywords and to predictable fields…","st":"tested","i":3,"c":5,"m":[["merge right_on",1.8],["argparse usage",1.4],["argparse description",1.1],["bare ArgumentParser",0.8],["merge on",0.8],["after prog=",0.3],["batch_size (control)",0.1],["read_csv sep (control)",-0.0],["figsize (control)",-0.0],["allow_redirects",-1.3]]},{"id":"L13_E199_N177","k":"m","L":13,"E":199,"N":177,"t":"A neuron that turns negative when 'things' get changed","l":"This neuron watches generic words for 'the situation': things, matters, всё, cosas. When such a word is about to be changed, as in 'make things worse', 'improve things' or 'to make matters worse', it fires strongly negative, down to about -6.6. The generic word is required…","st":"tested","i":5,"c":8,"m":[["control: economy",0.0],["control: dinner",0.0],["control: sense",0.0],["the thing",-1.5],["change things",-2.2],["regulate (halo+)",-2.5],["improve things",-2.8],["make matters",-5.1],["worsen things",-5.3],["make things",-5.9]]},{"id":"L13_E251_N16","k":"m","L":13,"E":251,"N":16,"t":"Repeats signed: mirror-closing symbols up, echoed prose words down","l":"This neuron sits in an expert whose router selects tokens that repeat something earlier in the text, so it reads almost only second mentions. It gives a signed verdict on each repeat. A symbol that closes a short mirror, such as the final 1 in 1 2 3 2 1, reads highest, and…","st":"tested","i":5,"c":7,"m":[["digit mirror",1.5],["rare fragment",1.0],["plain repeat",0.5],["adjacent repeat",0.1],["letter mirror",0.0],["word mirror",0.0],["options echo",0.0],["surname repeat",-0.0],["control: count",0.0],["control: letters",0.0]]},{"id":"L13_E37_N233","k":"m","L":13,"E":37,"N":233,"t":"Drawn out or holding several: a signed counting-membership neuron","l":"In counting problems, this neuron tracks which way membership runs. It fires positive when items are drawn out of a larger whole or when each member gets a single unit ('color some cells of a grid', 'each volunteer gets only one'). It fires negative, and more strongly, when one…","st":"tested","i":6,"c":6,"m":[["some cells of",2.9],["some sides of",1.6],["each only one",1.3],["two and one",0.9],["all cells of",0.3],["control: answer",0.0],["control: solve",-0.0],["two vars zero",-1.2],["contain two sides",-2.1],["more 9s",-4.3]]},{"id":"L13_E3_N35","k":"m","L":13,"E":3,"N":35,"t":"A South-Asia-routed neuron that reads head nouns, with signed exceptions","l":"This neuron sits in an expert the router reserves for South Asian and ancient history, plus a few memorized templates. The router picks the topic; the neuron reads the word. Inside that expert it fires positive on head nouns that close a phrase, most of all after a count…","st":"tested","i":3,"c":2,"m":[["region (core)",0.4],["hills",0.3],["coast (halo)",0.2],["kingdom",0.1],["campaign",-0.0],["was (control)",0.0],["and (control)",0.0],["revolt (neg)",-0.2],["islands (neg)",-0.3],["expedition (neg)",-0.5]]},{"id":"L13_E70_N183","k":"m","L":13,"E":70,"N":183,"t":"A neuron that remembers Annie Get Your Gun word pairs","l":"This neuron is a memorized-phrase lookup for Irving Berlin's Annie Get Your Gun. It fires on the second word of exact pairs such as 'Annie Get', 'Irving Berlin' and 'No Business Like', and goes quiet if the first word is changed even slightly. Silencing it makes the model much…","st":"tested","i":6,"c":7,"m":[["You Can Do",9.1],["No Business Like",8.5],["Annie Get",8.4],["Annie Oak",5.6],["Trade (control)",4.4],["They Can Do",4.3],["other title",0.9],["Anne (control)",0.5],["hymn verse",-1.3]]},{"id":"L14_E127_N311","k":"m","L":14,"E":127,"N":311,"t":"Closing school name up, invasions and market forces down","l":"Inside history-of-ideas prose, this neuron gives a signed reading. It goes positive on a school or style name that closes a list of peers, as in '…Taoist, Buddhist and Confucian', which is its strongest case. It goes negative on impersonal forces and upheavals such as invasions…","st":"tested","i":5,"c":5,"m":[["Confucian (core)",1.4],["Stoic",0.9],["Zoroastrian",0.2],["art style",-0.0],["set phrase",-0.0],["-ism piece",0.0],["goods control",0.0],["fruit control",0.0],["invasions",-0.3],["market forces",-0.3]]},{"id":"L14_E142_N82","k":"m","L":14,"E":142,"N":82,"t":"A neuron for 'but remember to…' after a friendly yes","l":"This neuron marks the moment a speaker gives someone permission and then adds a reminder: the 'be sure to', 'remember to' or 'don't forget' that follows 'of course you can'. It fires hardest on the first word of that care instruction. It fires about half as much on plain facts…","st":"tested","i":6,"c":5,"m":[["be sure",3.2],["remember",2.4],["don't forget",2.3],["be careful",2.3],["please",1.3],["forgetting (halo)",1.0],["control: number",0.3],["fact after but",0.1],["control: article",0.0],["public affairs",-0.5]]},{"id":"L14_E149_N331","k":"m","L":14,"E":149,"N":331,"t":"Spaces before or after the word: a table-padding sign neuron","l":"Inside text tables and boxes, this neuron reports where a run of spaces sits relative to the cell's word. Padding before the word, as on the first row under a header, drives it positive. Padding after the word, or a break that splits a word, drives it negative, and the negative…","st":"tested","i":3,"c":3,"m":[["plain prose",0.1],["long trail pad",0.1],["word, no pad",0.0],["clean row",0.0],["truncated word",0.0],["short lead pad",-0.1],["short trail pad",-0.1],["long lead pad",-0.2],["pad then word",-0.5]]},{"id":"L14_E173_N125","k":"m","L":14,"E":173,"N":125,"t":"A neuron that expects a just-set variable to be overwritten","l":"This neuron expects a variable to be overwritten. When an else, elif, except or guarding-if branch opens with the name of a variable that was just assigned, especially from an indexed read like right[i], it fires strongly. It does so because a fallback or clamp value is coming.…","st":"tested","i":7,"c":7,"m":[["same var",3.5],["after ' = 0'",3.5],["loop index",0.2],["pass",0.1],["print",0.1],["fresh var",-0.0],["other old var",-0.1],["opens while",-0.1],["opens for",-0.2]]},{"id":"L14_E183_N140","k":"m","L":14,"E":183,"N":140,"t":"A then-versus-now dial on car model names","l":"This neuron reads when a scene is set, and it reads it at the name of a product, above all a car. If earlier text places a Camry or CR-V in the past, through a 1980s-90s year, 'vintage' or 'back in 1985', the model-name token fires strongly. If the text places the product in the…","st":"tested","i":9,"c":8,"m":[["1985",5.8],["1972",4.9],["Reagan-era",4.1],["vintage",2.9],["1998",2.8],["body (control)",1.2],["color (control)",1.0],["2016",-0.6],["latest-gen",-1.3],["2025",-1.6]]},{"id":"L14_E191_N386","k":"m","L":14,"E":191,"N":386,"t":"A neuron for display formatters, quietly anti-correlated with game players","l":"This neuron responds to the word 'formatter' when it names a function that turns values into display text for a chart or UI component, scoring about 2.4 in a Tooltip sentence but near zero for a logging formatter. Weaker echoes come from other tool names and 'integer' used as an…","st":"tested","i":3,"c":5,"m":[["prefilter",1.1],["labelFormatter",1.0],["serializer",0.9],["formatter",0.0],["renderer",0.0],["number",0.0],["callback",0.0],["title",0.0],["color",0.0],["player",0.0]]},{"id":"L14_E196_N195","k":"m","L":14,"E":196,"N":195,"t":"A neuron that marks the first shortened re-mention of a coined name","l":"This neuron in Qwen3.6-35B-A3B fires negatively the first time a document shortens a name it coined itself. A full name like 'Edmund Vance Cooke' followed by a bare 'Cooke', a site title cut down in a footer, or an acronym for a phrase just defined all trigger it. Famous names…","st":"tested","i":7,"c":6,"m":[["exact repeat",-0.0],["other name",0.0],["no name",-0.0],["no antecedent",-0.0],["initial",-0.1],["distant mention",-1.4],["partial name",-1.6],["name + place",-1.7],["full name",-1.9]]},{"id":"L14_E19_N497","k":"m","L":14,"E":19,"N":497,"t":"A statistics neuron: setting up a test versus reporting results","l":"Inside statistics writing, this neuron separates setting up a test from reporting what was found. It fires on the word right after a test is named, and fires most when two groups are compared, as in 'tests of 2 means is'. It goes negative on verbs that present or attribute…","st":"tested","i":4,"c":5,"m":[["test of hypothesis",1.9],["test setup",1.7],["two-group test",0.7],["estimation halo",0.5],["discussing results",0.5],["interval halo",0.1],["neutral verb",0.1],["presenting results",0.1],["control: driving",0.0],["control: recipe",-0.0]]},{"id":"L14_E244_N439","k":"m","L":14,"E":244,"N":439,"t":"A rare-name lookup neuron: Soegaard up, 'century' down","l":"This neuron works like a short lookup table of word pieces, each with a fixed sign. It fires strongly on the middle piece of a few rare surnames, most of all 'ega' in Soegaard, and more weakly on rare scholarly terms. It goes negative on topical nouns such as 'century' in…","st":"tested","i":3,"c":6,"m":[["So|ega (core)",3.3],["Mo|ega (control)",0.5],["Sh|uji",0.4],["Zit|elli",0.3],["A|mano",0.2],["phylog|eny (halo)",0.0],["Takahashi (control)",-0.0],["function word",0.0],["century (neg)",-0.0],["Su|zy (neg)",-0.3]]},{"id":"L14_E248_N169","k":"m","L":14,"E":248,"N":169,"t":"A neuron for marked positions on a walker's path","l":"This neuron tracks a familiar move in programming-contest problems: listing which positions along a walker's path are marked, such as broken stairs, blocked grid cells, or sunk stones. It fires most strongly in a path setting with symbolic lists like a_1, ..., a_M. Subsets…","st":"tested","i":5,"c":8,"m":[["indexed broken",17.5],["literal numbers",11.4],["indexed painted",10.5],["grid cells",10.5],["unindexed",9.2],["distinct numbers",2.5],["different stones",1.2],["control: output",0.5],["control: prose",0.2]]},{"id":"L14_E55_N185","k":"m","L":14,"E":55,"N":185,"t":"A neuron for the 'what I'm seeking' slot in self-ads","l":"This neuron responds when someone presenting themself reaches the point where they say what they want. That point is the 'JOB OBJECTIVE' line of a résumé (about +2.7) or the 'seeking' / 'looking for' after a self-description in a personal ad. It is the slot that matters, not…","st":"tested","i":6,"c":7,"m":[["gay man",1.4],["job seeker",1.2],["woman, 34",1.1],["roommate-type",0.9],["SWM abbrev",0.8],["new grad",0.8],["employer",0.0],["employer 2",0.0],["narrative",-0.0]]},{"id":"L14_E68_N179","k":"m","L":14,"E":68,"N":179,"t":"A neuron that says 'encounter' in Korean and frowns at Japanese verbs","l":"This neuron marks the moment a Korean or Japanese verb begins after its marked noun, and the language sets the sign. It fires most strongly on the Korean root 접 ('come into contact with') after an object, as in 'encountering ancient artifacts'. Japanese particle+kanji verb…","st":"tested","i":4,"c":8,"m":[["KO 유물을 접",1.5],["KO 언어를 접",1.2],["JA 文章を構成",0.9],["KO 60을 얻",0.2],["control: English",0.1],["EN take",-0.0],["control: no verb",-0.0],["KO 능력을 가지",-0.2],["JA 母に見",-1.3],["JA clash",-3.9]]},{"id":"L14_E99_N77","k":"m","L":14,"E":99,"N":77,"t":"Spread wider than the spot, not dropped as a lump","l":"This neuron tracks how far a worked surface reaches around a spot. In instructions to sand, glue, paint or fill around damage, it fires on the words that set the zone's reach: 'an area larger than the patch', 'feather it out wide'. Size words with no surface work barely…","st":"tested","i":7,"c":7,"m":[["wider than",8.2],["feather wide",7.6],["larger than",7.2],["and surrounding",5.0],["smaller than",4.3],["entire",3.7],["dollop",2.0],["control: wait",0.7],["control: label",0.5],["blob",-1.1]]},{"id":"L15_E10_N83","k":"m","L":15,"E":10,"N":83,"t":"Where one part sits against another in a built figure","l":"This neuron marks the moment a text locates one part against a specific part of another, such as a circle sitting in a square's corner or a fin attached at a wing's end. It fires only after a figure or assembly has been set up, and the scene matters more than the spatial word.…","st":"tested","i":5,"c":5,"m":[["against edge",2.4],["in corner",1.7],["around center",1.7],["touches sides",0.0],["at vertex",0.0],["near middle",0.0],["color (control)",-0.0],["rank (control)",-0.0],["metric",-0.4]]},{"id":"L15_E137_N18","k":"m","L":15,"E":137,"N":18,"t":"A signed head-noun reader: doctrines up, stock compounds down","l":"This neuron reads the head noun that closes a phrase and gives it a sign. It turns positive when the noun names the tenets, plan or members of a system that the passage has already described, as in 'fundamental beliefs' in a church fact sheet. It turns negative when an everyday…","st":"tested","i":3,"c":6,"m":[["beliefs (core)",0.0],["strategy",-0.0],["arrangement",-0.0],["teachings (weak)",0.0],["list (halo)",-0.0],["control",0.0],["control",0.0],["stock compound",0.0],["stock compound",0.0],["stock compound",0.0]]},{"id":"L15_E144_N295","k":"m","L":15,"E":144,"N":295,"t":"A neuron for homes as the place wealth is stored","l":"This neuron goes strongly negative when text describes where a person's wealth is stored, and it goes deepest when that store is a home: 'net worth excluding their homes', 'non-liquid assets: primary residence'. The frame alone ('excluding their', 'tied up in her') already pulls…","st":"tested","i":7,"c":8,"m":[["herbs (control)",-0.0],["savings",-1.9],["stocks",-3.2],["business",-6.1],["jewelry",-6.5],["houses",-10.1],["cars",-10.3],["residence",-11.0],["homes",-14.4]]},{"id":"L15_E162_N6","k":"m","L":15,"E":162,"N":6,"t":"'Approach' after a problem goes up; a reply's opening quote goes down","l":"This neuron is a signed dial for segment openers. It turns positive on the word 'Approach' when it heads a method section right after a stated programming problem, peaking near +2 in code comments. It turns strongly negative on a bare quote mark that opens an assistant reply…","st":"tested","i":4,"c":6,"m":[["Approach",0.7],["Method",0.2],["Example",0.1],["Idea",-0.0],["Enfoque",0.0],["Solution Code",0.0],["References",-0.0],["quote mark",-0.0],["Key Insight",-0.1]]},{"id":"L15_E247_N204","k":"m","L":15,"E":247,"N":204,"t":"A neuron that flags the product after the sum","l":"This neuron in layer 15 of Qwen3.6-35B-A3B answers one math pattern with a strong negative signal: the value of a product given right after a sum, as in 'a + b = 1 and ab = −1'. It reaches about −5 there, and the router picks its expert first. Its weaker positive side marks…","st":"tested","i":6,"c":7,"m":[["control: prose",-0.0],["control: noun",-0.0],["ratio",-0.4],["second sum",-1.0],["a²+b²",-2.3],["xy",-3.2],["abc",-4.3],["ab (core)",-4.6],["a*b",-5.0]]},{"id":"L15_E251_N45","k":"m","L":15,"E":251,"N":45,"t":"A negative detector for capital unknowns inside formulas","l":"This neuron responds to capital-letter unknowns sitting inside formula factors. The X in (x - X) or in (X - a) pushes it strongly negative, and expectation or random-variable context deepens the dip to about -9. Lowercase letters barely register. Its positive side is about four…","st":"tested","i":5,"c":6,"m":[["self-composition",-0.0],["lowercase x",-0.0],["number",-0.0],["lowercase t",-0.5],["capital W",-2.0],["capital T",-2.3],["capital Y",-2.5],["capital A",-3.5],["capital X",-7.5]]},{"id":"L15_E56_N489","k":"m","L":15,"E":56,"N":489,"t":"A neuron for bridge expansion joints and the gaps they fill","l":"This neuron tracks one engineering idea: the gap left between bridge spans so they can expand and contract, and the hardware that fills it. It fires in many languages and hardest on the joint-type value in spec sheets ('expansion joint type: modular'), where zeroing it…","st":"tested","i":5,"c":6,"m":[["deck exp. joint",9.1],["gap seal",8.3],["exp. joint",7.3],["bearing",6.7],["girder",4.5],["median barrier",1.0],["railing",0.7],["kitchen cabinet",-0.0],["office furniture",-0.0]]},{"id":"L15_E63_N413","k":"m","L":15,"E":63,"N":413,"t":"Where the outtakes came from: a neuron for album sessions","l":"This neuron marks the stretch of time a creative work came out of, above all a record's recording sessions, and also its tour or era. It fires hardest when leftover material is traced back to those sessions, reaching about +12 on 'leftovers from the sessions'. Non-music working…","st":"tested","i":5,"c":6,"m":[["named sessions",8.2],["album sessions",7.1],["writing period",6.7],["album era",6.5],["committee (ctrl)",4.2],["tour",3.9],["holiday (ctrl)",1.3],["place (ctrl)",0.4],["credit 'for'",0.3]]},{"id":"L15_E91_N95","k":"m","L":15,"E":91,"N":95,"t":"A neuron that anticipates two opposites being joined","l":"This neuron watches for a familiar move in abstract writing: two complementary sides are named (reason and imagination, truth and grace, divine and human) and the text is about to say they belong together. It peaks on the pivot just before that union is stated, such as…","st":"tested","i":6,"c":8,"m":[["ideally",10.0],["preferably",8.2],["in practice",6.2],["unfortunately",3.0],["ultimately",2.2],["control: place",1.6],["historically",0.8],["control: day",0.6],["realistically",0.5]]},{"id":"L16_E100_N457","k":"m","L":16,"E":100,"N":457,"t":"A neuron that remembers one poem: Wang Wei's red beans","l":"This neuron recognizes a single memorized poem: Wang Wei's quatrain 相思 ('red beans grow in the south… these most stir longing'). Whatever brings the poem to mind fires it strongly, whether that is the title, any line, the red bean named as a love token, or the English title.…","st":"tested","i":8,"c":7,"m":[["red bean",13.4],["title",12.8],["first line",11.1],["last line",7.1],["English title",3.2],["君 verse",2.7],["Wang Wei other",2.2],["Li Bai",0.4],["Du Fu",0.3]]},{"id":"L16_E108_N212","k":"m","L":16,"E":108,"N":212,"t":"A neuron that goes negative when artworks become objects of study","l":"This neuron in Qwen3.6-35B-A3B goes sharply negative whenever text treats an artwork as something to study: art-history assignments, museum labels, catalog dates and art reading lists. Its strength depends on the art form. Paintings and buildings drive it hardest, music and film…","st":"tested","i":6,"c":10,"m":[["chemistry",-0.4],["recipe",-0.7],["novel",-0.8],["tax law",-1.1],["poem",-1.5],["symphony",-3.1],["film",-3.7],["building",-7.5],["sculpture",-10.7],["painting",-10.8]]},{"id":"L16_E131_N288","k":"m","L":16,"E":131,"N":288,"t":"A neuron that marks food as a source of illness","l":"This neuron gives a signed reading about food. Its strong side is negative: when text says food is moldy, toxic or teeming with microbes, as with aflatoxin on stored maize, Salmonella in pork or Chinese 霉变 'moldy' poisoning notices, it fires hard on the food nouns and fungus…","st":"tested","i":5,"c":6,"m":[["lead paint",-0.4],["undercooked chicken",-0.8],["pesticide rice",-0.8],["braised ribs",-0.9],["fresh peanuts",-1.2],["moldy peanuts",-2.5],["fresh sugarcane",-2.9],["Salmonella pork",-3.3],["moldy maize",-3.5],["moldy sugarcane",-4.7]]},{"id":"L16_E136_N47","k":"m","L":16,"E":136,"N":47,"t":"A neuron that knows you're about to say 'you're right'","l":"This neuron anticipates a reluctant concession. When someone has just argued, teased or corrected, and the reply begins to credit them ('I guess you're…', 'He did have…', 'I apologize, you're…'), it fires on the word just before 'right' or 'a point', whether or not that word…","st":"tested","i":8,"c":8,"m":[["core",7.9],["reluctant opener",5.7],["weak hedge",4.4],["generic assent",2.8],["third person",2.5],["plain opinion",1.9],["self subject",1.6],["control",0.0],["approval",-0.2]]},{"id":"L16_E141_N189","k":"m","L":16,"E":141,"N":189,"t":"A neuron for injury verbs and lifetimes, against haplogroups and calendars","l":"This neuron does not track one abstract idea. It fires on a small set of words that finish a phrase in a specific setting: 'banging' at the end of a list of ways to get hurt, 'cold air' as an asthma trigger, 'during his lifetime'. Its negative side is just as narrow. It dips…","st":"tested","i":2,"c":2,"m":[["scraping",0.2],["banging (core)",0.0],["punching",0.0],["scratching",-0.0],["twisting",-0.0],["noise verb",0.0],["any other way",-0.0],["control",-0.0],["control",-0.0]]},{"id":"L16_E161_N56","k":"m","L":16,"E":161,"N":56,"t":"A neuron for pronouns that don't fit the established person","l":"Once a story has set up one particular person, this neuron fires on a person pronoun that cannot refer back to them, most strongly a plural 'them' after a single named individual. Adding a compatible referent switches it off. Repeated names, matching pronouns and 'everyone'…","st":"tested","i":8,"c":6,"m":[["plural for one",2.8],["collective we",2.1],["new name",0.5],["group noun",0.4],["wrong gender",0.0],["control: object",0.0],["control: object",-0.0],["matching pronoun",-0.1],["quantifier",-0.1],["repeated name",-0.2]]},{"id":"L16_E171_N130","k":"m","L":16,"E":171,"N":130,"t":"Allowed letters versus four-byte characters a database cannot store","l":"This neuron tracks talk about which characters a system can handle, and the kind of character sets its sign. Naming an allowed letter or script class, such as non-ASCII, uppercase or Chinese, gives a weak positive push. Naming characters too wide for legacy storage, such as…","st":"tested","i":9,"c":6,"m":[["leading zeros",1.7],["Chinese",1.5],["letter case",1.0],["multibyte",0.6],["non-ASCII",0.5],["control",0.0],["two-byte",-0.4],["emoji",-2.2],["three-byte",-3.8],["four-byte",-8.7]]},{"id":"L16_E20_N342","k":"m","L":16,"E":20,"N":342,"t":"Deciding a fate versus distinguishing a kind: one signed neuron","l":"This neuron reads a signed axis on stakes predicates. It turns positive when a single event or factor settles an outcome (a battle that decided the fate of a kingdom, a decisive vote, life-or-death stakes) and negative, more strongly, when a trait is named as what distinguishes…","st":"tested","i":5,"c":6,"m":[["deciding factor",1.5],["decided fate",1.3],["settled fate",1.1],["turning point",0.4],["control: increased",0.2],["decisive",0.0],["control: measured",-0.3],["defined level",-0.4],["distinguished from",-1.8],["separates from",-3.0]]},{"id":"L16_E226_N144","k":"m","L":16,"E":226,"N":144,"t":"Germs and particles versus the bodies that carry them","l":"This neuron tracks loose, tiny matter in relation to something it sits on, enters or comes off. It is positive on germs, spores, toner, dust and loose yeast, and strongest when the text is about contamination or cleaning. It turns negative on the larger body that carries or…","st":"tested","i":6,"c":8,"m":[["spores",1.1],["bacteria",0.9],["dust",0.9],["germs",0.7],["crumbs (halo)",0.4],["control: person",-0.0],["control: plan",-0.2],["pets",-1.1],["infected host",-2.2]]},{"id":"L16_E24_N383","k":"m","L":16,"E":24,"N":383,"t":"A neuron that fires where an announced induction proof begins","l":"This neuron reads a plan rather than a word. It fires where a math or algorithm answer that has just announced induction opens its base step, on the token that starts the proof, such as '1.' or 'First'. Recurrence seeds, dynamic-programming base cases and small-case checks get a…","st":"tested","i":8,"c":6,"m":[["induction on n",3.0],["strong induction",2.9],["induction",2.2],["DP",0.5],["recursion",0.5],["contradiction",-0.2],["control: formula",-0.2],["control: tool",-0.2],["greedy",-0.4],["L'Hôpital",-0.5]]},{"id":"L16_E40_N449","k":"m","L":16,"E":40,"N":449,"t":"A detector for pronounceable-looking consonant runs without vowels","l":"This neuron fires on the second or later piece of a letter string that has gone without a vowel and is built from common English consonants: disemvoweled handles, repo slugs, random passwords, keysmash. A handle like @pltzkrm drives it to about +5, while an equally long string…","st":"tested","i":5,"c":8,"m":[["core slug",5.9],["core run 2",4.8],["core run",4.5],["one rare letter",1.1],["digits mixed",1.0],["rare letters",0.2],["vowelled control",0.0],["vowel resets",0.0],["word control",-0.1]]},{"id":"L16_E54_N167","k":"m","L":16,"E":54,"N":167,"t":"Same verb, opposite sign: a role's effect versus a target","l":"This neuron reads the verb that follows 'to', 'not' or 'and', and the surrounding frame decides its sign. If the text names what something acts or functions as, the verb stating what it accomplishes fires positive, as in zinc that 'does not allow' water through. If the same…","st":"tested","i":5,"c":6,"m":[["role (acts as)",1.3],["goal",0.5],["serves as",0.3],["designed to",0.1],["function as",-0.0],["plain effort",-0.0],["control: negation",0.0],["control: adverb",-0.0],["how-to",-0.0],["contest",-0.2]]},{"id":"L16_E55_N220","k":"m","L":16,"E":55,"N":220,"t":"A neuron for things that can still change","l":"This neuron fires when text says something is not yet fixed and can still be altered. A schedule, forecast, ranking or just 'things' may still change, be revised or be added to. 'Still' or 'yet' strengthens it sharply, and it works across languages. Time left before a deadline…","st":"tested","i":6,"c":7,"m":[["can still change",8.2],["no 'still'",7.1],["negated",7.0],["still revised",6.5],["non-alteration",6.3],["subject to",5.8],["improvement",3.9],["indefinite other",1.1],["past control",0.0],["settled control",0.0]]},{"id":"L16_E60_N479","k":"m","L":16,"E":60,"N":479,"t":"A neuron that remembers single words in specific memorized passages","l":"This neuron behaves less like a concept detector than like a small set of memory entries. It fires on one particular word inside one particular memorized passage: the 'if' of Sétif in the 1945 massacre sentence, 'reader' in Kubernetes metrics-reader configs, or 'integer' in the…","st":"tested","i":3,"c":2,"m":[["wrong massacre",0.1],["Guelma",0.0],["Kherrata",0.0],["Constantine",0.0],["Algiers",0.0],["Oran",-0.0],["Paris control",0.0],["wrong massacre 2",0.0],["Sétif (memorized)",-1.8],["Sétif+Guelma",-1.8]]},{"id":"L16_E75_N149","k":"m","L":16,"E":75,"N":149,"t":"A neuron for the expected word that completes familiar text","l":"This neuron responds to the word that completes familiar wording: the closing half of a sutra pair, 'build upon' at the end of a licence's verb list, or 'Precipitation is light' after a climate description. Context builds the response, so the same phrase on its own barely…","st":"tested","i":4,"c":6,"m":[["light (core)",-0.0],["concentrated",0.0],["spread",-0.0],["sparse",0.0],["heavy (halo)",-0.0],["moderate",-0.0],["purple (control)",0.0],["Tuesday (control)",-0.0],["5px (negative)",0.0]]},{"id":"L16_E85_N346","k":"m","L":16,"E":85,"N":346,"t":"Qualities you must have versus certificates you must submit","l":"This neuron reads requirement and eligibility text. It fires positive on the condition a person must meet to qualify, such as physical preparation, solid knowledge, resembling an ID photo, or a car being in its place. It fires negative on certified records that must be…","st":"tested","i":5,"c":6,"m":[["preparation",2.7],["knowledge",1.7],["condition",1.3],["funds",0.7],["weekend (control)",0.5],["password",0.4],["umbrella (control)",0.3],["passport (control)",0.2],["birth cert.",-1.4],["incorporation",-2.6]]},{"id":"L17_E114_N163","k":"m","L":17,"E":114,"N":163,"t":"A neuron for what's left out: meat, caffeine, the state","l":"This neuron in the middle of Qwen3.6-35B-A3B marks the item that something is defined by lacking: meat in a vegetarian diet, direct identifiers in pseudonymized data, the state for anarchists. Items that do not fit the frame stay silent. Its negative side is the opposite…","st":"tested","i":6,"c":4,"m":[["fish",1.2],["poultry",0.8],["eggs",0.7],["dairy",0.6],["milk (halo)",0.4],["sugar",0.1],["meat (core)",0.0],["shoes (control)",-0.0],["Mondays (control)",-0.0],["music (control)",-0.0]]},{"id":"L17_E115_N248","k":"m","L":17,"E":115,"N":248,"t":"Open big questions push it up; specific case questions push it down","l":"This neuron fires at question boundaries and its sign tracks what kind of question is being asked. A run of open conceptual questions (is X fair, does X exist, what is X) drives it up, and each new question adds more. Determinate questions about a specific case, such as a math…","st":"tested","i":7,"c":7,"m":[["mind question",3.8],["math",3.6],["mechanism",3.5],["crime scene",3.5],["aesthetics",2.8],["ethics",2.8],["errand",2.8],["history",2.7],["control: request",0.3],["control: statement",-0.4]]},{"id":"L17_E15_N359","k":"m","L":17,"E":15,"N":359,"t":"A neuron that marks the first Sample Input heading","l":"This neuron marks the moment a programming-contest problem statement turns to its first worked example: it peaks on the word 'Input' in the first 'Sample Input' heading and responds more weakly to other headings in the statement. It turns negative on 'nominations' at the end of…","st":"tested","i":4,"c":4,"m":[["Sample Input",1.0],["Sample Output",0.9],["入力例",0.4],["Input example",0.4],["Constraints",0.3],["Examples",0.2],["prose control",-0.0],["note control",0.0],["boxed answer",-0.6],["nominations",-1.1]]},{"id":"L17_E162_N436","k":"m","L":17,"E":162,"N":436,"t":"Building a unit of size 3 versus picking 3 from a set","l":"This neuron reads what job a small number like 2 or 3 is doing. It fires positive when the number sets the size of one unit being built along a sequence: a three-character window, a pair to form, a loop that starts at 2, pages per sheet. It goes negative when the same number…","st":"tested","i":5,"c":7,"m":[["subset ≥N",4.2],["first N",3.6],["run",3.3],["triple",3.2],["window",1.9],["control",1.8],["cartons",1.7],["control",1.0],["k-th largest",0.7]]},{"id":"L17_E172_N235","k":"m","L":17,"E":172,"N":235,"t":"A period that marks 'here is what this code does'","l":"This neuron fires on the punctuation that closes a sentence explaining what a just-introduced software, web or math component does: 'This file defines…', 'In this solution, the Node struct is used to represent…'. The same frame applied to hinges, paintings or recipes stays flat.…","st":"tested","i":4,"c":5,"m":[["code gloss",1.3],["plugin purpose",0.7],["file purpose",0.2],["heraldry",0.1],["math result",0.0],["physical halo",0.0],["hinge (control)",0.0],["cooking (control)",0.0],["fossil anatomy",-0.1],["math dead end",-0.6]]},{"id":"L17_E178_N193","k":"m","L":17,"E":178,"N":193,"t":"A signed list detector: activity lists positive, course catalogs negative","l":"This neuron tracks what kind of list is being written. It is most positive where a list of activities or skills begins, often at the very first bullet after 'learn to:', before any item appears. It turns negative inside dry lists of academic topics in course catalogs and in…","st":"tested","i":5,"c":7,"m":[["kids enjoy",1.4],["learn to",1.3],["capable of",0.8],["math catalog",0.5],["job duties",0.0],["topics (noun)",0.0],["features (noun)",-0.0],["ingredients",0.0],["notes (control)",0.0],["side effects",0.0]]},{"id":"L17_E184_N354","k":"m","L":17,"E":184,"N":354,"t":"A neuron for fresh symbols that stand for anything","l":"This neuron fires on a symbol at the moment it is introduced to stand for any member of a set, rather than to name one thing. Examples are M after nA nB nC nD, the X in HgX2 for any halide, and a fresh x after 'for any integer'. Letters that already name something, such as a…","st":"tested","i":7,"c":8,"m":[["fresh c",5.4],["fresh x",4.1],["fresh m",1.9],["word control",0.0],["bound b",-0.1],["fixed number",-0.1],["constant",-0.1],["index n",-0.2],["bound a",-0.3],["named sequence",-0.3]]},{"id":"L17_E212_N175","k":"m","L":17,"E":212,"N":175,"t":"A neuron that signs pre-action phrases: means positive, premise negative","l":"This neuron watches the slot just before an action verb. When the phrase there says how the action will be done (a tool, a heat level, a guide, a format such as 以UTF-8编码, or a rotation axis), it fires positive. When the phrase instead cites a premise or condition, as in…","st":"tested","i":6,"c":8,"m":[["tool",1.3],["guide",1.2],["heat setting",0.5],["premise",0.5],["weak tool",0.4],["workspace",-0.0],["control",0.0],["noun control",-0.0],["helpers",-0.4],["resource limit",-0.4]]},{"id":"L17_E215_N296","k":"m","L":17,"E":215,"N":296,"t":"A neuron for helium listed right after neon","l":"This neuron fires on the lightest member of a chemical family, such as helium or lithium, when it appears in a list right after a heavier sibling: 'neon, helium' fires strongly, but 'helium, neon' barely does. It works across languages and symbols. Bracket and operator partners…","st":"tested","i":5,"c":4,"m":[["helium",4.0],["He symbol",1.9],["lithium",1.1],["hydrogen",0.4],["heavy noble gas",0.1],["heavy noble gas",0.1],["French helium",0.0],["control",-0.0],["non-element",0.0],["sodium",-0.2]]},{"id":"L17_E247_N47","k":"m","L":17,"E":247,"N":47,"t":"Setting up a quantity versus concluding from the evidence","l":"This neuron tracks two kinds of moment in quantitative text. It turns positive when a problem or explainer names the quantity it is about to reason over, as in 'for every positive integer' or 'pay semiannual interest'. The word and the frame must both be right. It turns strongly…","st":"tested","i":4,"c":5,"m":[["integer (core)",1.7],["divisor",0.6],["reason to believe",0.2],["finance setup",0.0],["premise (halo)",0.0],["control",-0.1],["control",-0.1],["sign-off",-0.1],["warrant",-0.1],["rate (halo)",-0.2]]},{"id":"L17_E33_N266","k":"m","L":17,"E":33,"N":266,"t":"A recall marker for a handful of verbatim web passages","l":"This neuron acts as a recall marker for a small, fixed set of passages that recur word for word on the web. It fires inside the Passover 'unleavened bread' verse, contest-math constants like f(1) = -2016, the Wikipedia footer and stock word-problem templates. A different set of…","st":"tested","i":3,"c":6,"m":[["near variant",2.8],["exact verse",2.7],["paraphrase",2.6],["synonym",2.4],["control",2.4],["halo",2.3],["verse phrase",2.2],["control",1.4],["control",1.3]]},{"id":"L17_E37_N173","k":"m","L":17,"E":37,"N":173,"t":"A detector for words the source never said","l":"This neuron flags the moment a reply that is supposed to follow a source says something the source never gave. A translation that adds 'white' to an apron, a reconstructed premise that invents a golden wheat field, or an answer of 'Option A' when the options start at B all fire…","st":"tested","i":10,"c":7,"m":[["added color",4.3],["added adjective",2.2],["two additions",1.8],["added item",1.1],["substitution",1.0],["faithful + period",0.3],["near synonym",0.2],["faithful",0.1],["verbatim copy",0.1]]},{"id":"L17_E49_N456","k":"m","L":17,"E":49,"N":456,"t":"Participles standing alone as labels, not as verbs","l":"This neuron fires on the final piece of a participle (-ed, -ing) when the word stands on its own as a descriptor: a status value, an answer, a tag, or the start of a phrase. The response is strongest when the tokenizer has split the word. The same word used as a verb ('was…","st":"tested","i":4,"c":6,"m":[["split participle",2.7],["-ing",2.6],["single-token -ed",2.1],["split participle",1.7],["split -ed",1.3],["rare split",1.2],["-able halo",0.5],["abstract noun",0.2],["control",0.2],["control",0.0]]},{"id":"L17_E4_N341","k":"m","L":17,"E":4,"N":341,"t":"A neuron for members of textbook disease and hormone schemas","l":"This neuron responds to members of a few heavily taught medical stories once the story is under way. Examples are acid and pepsin in peptic ulcer, angiotensin in the renin-angiotensin-aldosterone chain, pituitary in hypothalamic-pituitary axes, and CSF protein in Guillain-Barré.…","st":"tested","i":4,"c":6,"m":[["pepsin (core)",4.7],["trypsin (halo)",2.1],["bicarbonate",1.0],["bile",1.0],["control",0.7],["wrong schema",0.2],["mucus",0.1],["control",0.0],["monomer name",-0.3],["chemical name",-0.6]]},{"id":"L17_E50_N382","k":"m","L":17,"E":50,"N":382,"t":"A neuron for gut feelings and premonitions of what's coming","l":"This neuron fires when a text describes someone knowing what is coming through an inner faculty: a gut feeling, a presentiment, 'something told me', 'can't shake this feeling'. The source has to be inside the person. 'My gut told me' scores about 6, while 'my lawyer told me'…","st":"tested","i":8,"c":8,"m":[["something",6.5],["dream",4.2],["gut",4.0],["presentiment",3.3],["heart",2.9],["lawyer",1.5],["warning",1.0],["report",0.8],["control",0.8]]},{"id":"L17_E65_N201","k":"m","L":17,"E":65,"N":201,"t":"Math text: conditions being stated versus numbers being combined","l":"In olympiad and programming text, this neuron acts as a signed axis. It goes positive while a requirement is being stated, such as 'every integer can be written in the form', 'the interval [', or 'all edges must go from'. It goes negative on the connecting word of an arithmetic…","st":"tested","i":5,"c":6,"m":[["condition",0.4],["static geometry",0.3],["relational (control)",0.3],["condition variant",0.2],["non-math (control)",0.1],["imperative geometry",0.0],["move (control)",0.0],["set removal",-0.1],["arithmetic",-0.1],["equation elimination",-0.6]]},{"id":"L17_E75_N399","k":"m","L":17,"E":75,"N":399,"t":"Ratios to be stated versus equations solved only numerically","l":"This neuron in Qwen3.6-35B-A3B tracks whether technical text is about to state what a ratio is or what it equals. It fires positive when a dimensionless coefficient such as Cp is about to be defined, or when a ratio's value is about to be given. It turns negative for equations…","st":"tested","i":4,"c":4,"m":[["dimensionless coef.",1.3],["LaTeX ratio",0.9],["cubic root",0.8],["ODE shooting",0.3],["ratio squared",0.2],["fraction arithmetic",0.0],["area (control)",0.0],["length (control)",0.0],["volume (control)",0.0],["tan fixed point",-3.4]]},{"id":"L17_E75_N421","k":"m","L":17,"E":75,"N":421,"t":"A neuron that marks the running parameter of a limit","l":"In math reasoning, this neuron marks the symbol that carries a limit or indexes a family of curves: the t set equal to a variable heading to infinity, or the ε defined as a small parameter. It fires about seven times as strongly when a limit is stated nearby. Words like 'small'…","st":"tested","i":6,"c":6,"m":[["a → 0",3.9],["for large a",3.1],["large/small",3.0],["a → ∞",2.3],["line family",1.7],["no limit",0.6],["held fixed",0.5],["single circle",-0.1],["non-math",-0.2]]},{"id":"L17_E86_N175","k":"m","L":17,"E":86,"N":175,"t":"Familiar strings push it up, loop bounds and accumulators push it down","l":"This neuron acts as a signed dial on a small set of digits and code identifiers that the router sends it. It reads positive when a token completes a familiar, memorized string, such as the year 1603, 'import sys', or 'stop' after 'start'. It reads negative when the token sits…","st":"tested","i":4,"c":6,"m":[["1603 (famous)",1.4],["1803",0.8],["1833",0.1],["1763",-0.0],["1663",-0.0],["1903",-0.0],["1403 (obscure)",-0.0],["no year",-0.0],["vague time",0.0],["non-date",0.0]]},{"id":"L18_E12_N146","k":"m","L":18,"E":12,"N":146,"t":"A neuron for my child's fear of death","l":"This neuron tracks fear in a particular person. It swings strongly negative when someone close to the writer, especially their own child, is afraid of something. The deepest values come for fear of death, then the dark and monsters. Its positive side is much weaker…","st":"tested","i":7,"c":7,"m":[["anticipation",0.8],["control: sport",-0.0],["control: food",-0.3],["worry",-2.4],["fear of dentist",-3.9],["fear of dark",-4.3],["death, no fear",-5.0],["fear of future",-6.4],["terror of dying",-9.5],["fear of death",-11.5]]},{"id":"L18_E13_N297","k":"m","L":18,"E":13,"N":297,"t":"A negative signal for starting to solve a word problem","l":"This neuron fires almost only in one direction: negative. It goes down when a text has stated a problem with quantities and then begins to solve it. 'First, let's' opens the state, and the deepest dip comes at the slot just before the number after '=' or 'divided by', near -7.7.…","st":"tested","i":5,"c":8,"m":[["Second,",0.1],["To start",-0.0],["no comma",-0.0],["narrative",-0.0],["Then,",-0.3],["bare Let's",-1.7],["First, we",-4.0],["First, let's",-4.4],["Firstly, let's",-4.5]]},{"id":"L18_E150_N73","k":"m","L":18,"E":150,"N":73,"t":"A neuron for what a substance travels inside when it crosses a boundary","l":"This neuron tracks bookkeeping of what crosses a boundary inside something else. It fires on the preposition and carrier word when a component rides out of or into a system within a medium: entropy crossing with the heat, cadmium excreted in the urine, nitrogen leaving in sold…","st":"tested","i":6,"c":7,"m":[["entropy",5.6],["phosphorus",2.5],["contaminant",2.4],["sediment (halo)",2.2],["nutrient",2.1],["merchandise",1.8],["control: vehicle",1.5],["bulk good",1.4],["cargo",1.0],["control: person",-0.1]]},{"id":"L18_E153_N301","k":"m","L":18,"E":153,"N":301,"t":"A neuron for 'the workmanship is exquisite'-style judgments","l":"This neuron fires when Chinese text states the value of a named quality dimension, as in 制作考究 ('the workmanship is exquisite'): a craft, manner or style noun is followed directly by a literary adjective. Plainer adjectives, inserted degree words, price-like dimensions and…","st":"tested","i":5,"c":8,"m":[["craft + literary adj",2.8],["style variant",2.8],["craft variant",2.4],["plain adj",1.8],["with 很",1.5],["no aspect noun",1.1],["price dimension",1.0],["control: fact",0.0],["narrative manner",-0.1],["control: count",-0.1]]},{"id":"L18_E174_N161","k":"m","L":18,"E":174,"N":161,"t":"A clause-head neuron: his cousin positive, the receiver negative","l":"This neuron sits in an expert that the model mostly uses inside sentence-opening framing clauses ('Since...', 'According to...', 'During...'). Within those clauses it marks the clause's head word. The value is positive when a possessive ties the head to a person's relative or…","st":"tested","i":4,"c":6,"m":[["his cousin",0.9],["his surname",0.7],["his kinsman",0.6],["his student",0.6],["control: house",0.5],["his report",0.1],["this review",-0.0],["source name",0.0],["the receiver",-0.0],["control: weather",0.0]]},{"id":"L18_E192_N494","k":"m","L":18,"E":192,"N":494,"t":"An 'and' that knows whether you asked for examples or definitions","l":"This neuron watches the connective that joins two parts of a request or outline, and its sign records what the first part asked for. After a request for particular instances ('Name 4 examples of botnets and') it turns positive. After a request for what something is ('What is X…","st":"tested","i":7,"c":8,"m":[["examples",2.4],["famous",1.9],["kinds",1.3],["history",0.3],["control: narrative",-0.0],["control: news",-0.0],["advantages",-0.2],["meaning",-2.0],["what are",-2.3],["definition",-2.6]]},{"id":"L18_E198_N75","k":"m","L":18,"E":198,"N":75,"t":"Clean camera settings versus the cramped, dark places you shoot","l":"Inside camera-exposure text, this neuron signs a single contrast. It turns positive when the text names the clean result or the base setting, such as 'ISO 100' or 'image quality', and it does the same for amplifier manuals that ask for a signal 'within the range'. It turns…","st":"tested","i":5,"c":7,"m":[["image quality",5.0],["cleanest quality",3.6],["audio halo",3.5],["sharp detail",2.4],["neutral list",0.9],["control: bread",0.1],["control: desks",-0.6],["dark venues",-2.6],["tight spaces",-3.4],["cramped rooms",-5.8]]},{"id":"L18_E19_N66","k":"m","L":18,"E":19,"N":66,"t":"A neuron for stated size: '8 characters', '3 meters', 'twice'","l":"This neuron fires on the word that finishes a statement of how big something is, right before 'long', 'tall' or 'as'. Examples are 'characters' in 'the password must be at least 8 characters', 'meters' in 'a wall 3 meters', and 'times' in 'the planet is 1.4 times'. The same…","st":"tested","i":5,"c":7,"m":[["characters",8.9],["bytes",5.2],["chars",5.0],["digits",4.5],["words",2.6],["letters",2.1],["dollars (price)",1.1],["hours",0.4],["users (count)",0.2],["percent",0.1]]},{"id":"L18_E203_N347","k":"m","L":18,"E":203,"N":347,"t":"A neuron that separates the poster's edits from the responder's handling","l":"This neuron tracks who is speaking about a posed task. It fires when the person who posted a problem or set up a game amends it: an 'Edit:' note under the problem, a disclaimer about what they pasted, or rules they will enforce. It goes negative when the text is about the…","st":"tested","i":6,"c":6,"m":[["off-topic edit",2.1],["edit note",1.9],["P.S. disclaimer",1.9],["requester rule",0.6],["control prose",0.6],["user aside",0.5],["control math",0.2],["answer opener",-0.7],["answerer assumption",-1.3]]},{"id":"L18_E224_N378","k":"m","L":18,"E":224,"N":378,"t":"Submitted work versus callbacks: a neuron that reads block hand-offs","l":"This neuron reacts to the moment code hands a block of work to another API, and its sign depends on what that API will do with the block. It fires most on Swift's DispatchQueue ... async {, where the block is submitted to run on a queue. It goes strongly negative when the block…","st":"tested","i":7,"c":6,"m":[["global async",2.8],["queue async",2.3],["new thread",2.3],["addOperation",1.4],["dataTask",1.0],["func (control)",0.4],["if (control)",0.0],["forEach",-1.0],["addObserver",-1.8]]},{"id":"L18_E227_N236","k":"m","L":18,"E":227,"N":236,"t":"A code neuron for adding children to UI containers","l":"Inside code, this neuron marks one kind of action: adding a child to a UI container. It goes strongly negative on telebot keyboard .add, table row.insertCell and stack-view addArrangedSubview, and only weakly negative on ordinary list or set appends. On the positive side it…","st":"tested","i":4,"c":5,"m":[["tk.Tk",0.5],["tkinter grid",0.1],["Qt addWidget",0.1],["googleapiclient",-0.0],["loop 'in'",0.0],["control: print",-0.0],["control: math",-0.0],["set.add",0.0],["markup.row",-1.1],["markup.add",-1.3]]},{"id":"L18_E244_N232","k":"m","L":18,"E":244,"N":232,"t":"A neuron that fires when a study's first result is due","l":"This neuron watches for the moment a study has been fully described but its result has not yet been given. It fires on the word that opens the result, such as 'As', 'The results showed' or 'found that'. More study details make it fire harder, and it goes quiet once a result has…","st":"tested","i":7,"c":8,"m":[["As (core)",4.9],["In",2.7],["Thus",1.8],["The results",1.6],["The (halo)",1.6],["Compared",0.6],["Liver (control)",0.1],["Serum (control)",-0.0],["Statistical (control)",-0.1],["more methods",-0.2]]},{"id":"L18_E36_N181","k":"m","L":18,"E":36,"N":181,"t":"A neuron for glued-on pollen packets, opposed to host plants","l":"This neuron tracks a peculiar piece of plant engineering: a packet the plant makes that carries its own glue and sticks onto an animal. Orchid and milkweed pollinia on a sticky viscidium are the clearest case. Both the idea and the word 'viscidium' drive it, and together they…","st":"tested","i":7,"c":8,"m":[["viscidium",10.9],["viscin seed",8.2],["milkweed",6.0],["concept only",5.3],["vectored spores",2.8],["generic sticky",0.7],["loose pollen",0.4],["orchid, no pollinia",-0.3],["fungal host",-8.8],["larval host",-8.9]]},{"id":"L18_E41_N269","k":"m","L":18,"E":41,"N":269,"t":"A neuron for national surplus about to be sent abroad","l":"This neuron tracks one storyline: a country produces more goods, capital or men than home can absorb, and the text is about to send that surplus abroad. It peaks on the connective right before the outlet is named. States sending troops or envoys outward fire almost as strongly.…","st":"tested","i":6,"c":6,"m":[["overcapacity",4.4],["national surplus",3.8],["surplus capital",3.5],["surplus troops",1.6],["shortfall",0.6],["local surplus",-0.0],["garden control",-0.0],["personal trip",-0.0],["code control",-0.0]]},{"id":"L18_E58_N360","k":"m","L":18,"E":58,"N":360,"t":"One empty operand decides it, or every operand must run out","l":"This neuron reads compound emptiness tests in code and gives them a sign according to which operand can decide the result. It fires positive when a single empty operand settles the condition: 'if a == 0 or b == 0', 'either child is null', 'while both pointers remain'. It swings…","st":"tested","i":9,"c":8,"m":[["any zero",7.2],["either null",5.9],["both remain",4.3],["plain constant",3.3],["partial bound",0.9],["control: call",0.0],["control: sum",-0.4],["any flag",-1.9],["carry loop",-4.6],["all None",-5.8]]},{"id":"L18_E85_N81","k":"m","L":18,"E":85,"N":81,"t":"A neuron for qualifiers in repeated records, flipped negative by Java throws","l":"This neuron fires on the small qualifier that closes a value in a repeated structured record. Its main cases are a flag that marks a field as mandatory (required=True, NOT NULL), a footnote mark on a table cell, and the closing token of a modular-arithmetic line. The same open…","st":"tested","i":3,"c":5,"m":[["required=True",2.1],["after dest",2.0],["spaced variant",1.8],["negated",0.8],["default (halo)",0.7],["action control",0.3],["type control",0.2],["help control",0.1]]},{"id":"L19_E105_N315","k":"m","L":19,"E":105,"N":315,"t":"A neuron for integrative structural biology and crystallography cell keys","l":"Inside a mixture-of-experts language model, one neuron fires almost only in protein-structure text. Its two strongest triggers are the word 'integrative' after several experimental methods have been combined into one model, and the '_cell' key in crystallography data files. It…","st":"tested","i":4,"c":4,"m":[["integrative",5.0],["hybrid",3.0],["combined",1.2],["mol. replacement",0.9],["histogram (ctrl)",0.3],["AlphaFold (halo)",0.2],["spreadsheets (ctrl)",0.2],["coffee (ctrl)",0.0],["IMOD",-0.6],["subtomogram",-1.3]]},{"id":"L19_E137_N294","k":"m","L":19,"E":137,"N":294,"t":"Putting a named theory to work in technical prose","l":"This neuron marks the moment when technical writing takes a theory, method or mechanism it has already named and commits it to a purpose, as in 'applied the theory to prove' or 'this is how the bulb makes light'. Its opposite pole is uses that are only claimed or advertised…","st":"tested","i":5,"c":6,"m":[["applied",2.6],["used",2.2],["employed",1.2],["adapted",0.9],["recipient",0.8],["wrote about",0.7],["advertised",0.7],["sent",0.2],["place",0.0]]},{"id":"L19_E149_N224","k":"m","L":19,"E":149,"N":224,"t":"A neuron that fires when algebra recovers the original variable","l":"This neuron tracks one move in step-by-step algebra: getting back an original variable after it was repackaged into new symbols. When a derivation defines s = b - d and t = b + d and then writes 't - s = 2d', the neuron fires on that final d. It ignores the new names and fresh…","st":"tested","i":9,"c":5,"m":[["bare original",8.1],["after plus",8.1],["original var",7.8],["exponent piece",6.0],["wrong original",3.8],["new name bare",0.1],["fresh symbol",0.0],["number",0.0],["new name",-0.1]]},{"id":"L19_E159_N368","k":"m","L":19,"E":159,"N":368,"t":"A neuron for sewing the patient back up","l":"This neuron tracks the final stage of an operation: after the work inside a body is done, the text names the incision or flap and closes it, or says it stays open. Surgical and veterinary notes drive it hardest. Repairs that reopen walls, cases or trenches give a weaker echo…","st":"tested","i":6,"c":8,"m":[["drywall",6.4],["surgery",5.2],["laptop",3.8],["dental flap",3.5],["suitcase",2.9],["surgery 2",2.0],["autopsy",0.7],["tire change",0.5],["spreadsheet",0.0],["meeting",0.0]]},{"id":"L19_E163_N27","k":"m","L":19,"E":163,"N":27,"t":"Statement-end neuron: demo-script code positive, algorithm index setup negative","l":"This neuron fires at the ends of code statements and takes its sign from the kind of code around them. Flat tutorial lines built from constants, such as a list of torch.zeros(3, 4) calls, push it positive, and the value grows as more of these lines appear. Bookkeeping lines…","st":"tested","i":4,"c":6,"m":[["torch run",0.9],["one torch line",0.8],["colour tuple",0.1],["variable call",0.1],["empty function",0.0],["comment",-0.0],["layer literal",-0.1],["index offsets",-2.3],["path join",-2.4]]},{"id":"L19_E193_N111","k":"m","L":19,"E":193,"N":111,"t":"A neuron for tax withheld at source, opposed by fines","l":"This neuron tracks one administrative frame: tax taken from someone's pay or investment income by an employer, broker or tax agent. It fires just before the withholding party is named. Neither the verb 'withheld' alone nor tax alone is enough. Visa sponsors and non-tax wage…","st":"tested","i":5,"c":6,"m":[["tax withheld",9.1],["tax collected",6.4],["visa sponsor",6.1],["dividend TDS",5.9],["pension",4.7],["union dues",1.2],["sales tax",1.1],["control: evidence",0.3],["control: bill",-0.0],["fine",-1.2]]},{"id":"L19_E201_N164","k":"m","L":19,"E":201,"N":164,"t":"A neuron for breaking ties: 'if both are plausible, pick'","l":"This neuron marks the moment a text settles an open choice. It fires hardest on 'pick' after an instruction admits that both options may be plausible, and on 'Maybe the' or 'I will assume' after a math solver notes that the data are missing. Giving up out of frustration and…","st":"tested","i":5,"c":6,"m":[["multi-correct",1.2],["both plausible",1.1],["equal tie",1.1],["unsure",1.0],["plain rule",0.8],["clear-cut",0.5],["give-up",0.5],["control",0.0],["promo",-0.5]]},{"id":"L19_E203_N67","k":"m","L":19,"E":203,"N":67,"t":"Login markup fires positive, sign-up steps swing negative","l":"This neuron lives inside text about user accounts and logins, and the role of each token decides its sign. On the machinery of authentication it fires positive: the tag that opens a field label in a login form, or config keys like an identity provider's authority or a registry…","st":"tested","i":5,"c":8,"m":[["login",2.3],["registration",1.6],["reset pw",0.8],["login (es)",0.5],["recipes (ctrl)",0.0],["contact (ctrl)",-0.5],["newsletter (ctrl)",-0.5],["create account",-1.1],["sign in",-1.6]]},{"id":"L19_E221_N174","k":"m","L":19,"E":221,"N":174,"t":"A neuron for 'trying to change', not 'trying to convert'","l":"This neuron marks the verb in a statement of purpose or attempt ('a framework to guide', 'trying to change'), but only when the verb means altering or steering something. In the same sentence 'change' fires strongly while 'convert' barely registers. Its opposite pole is the…","st":"tested","i":5,"c":6,"m":[["change",1.1],["modify",0.8],["fix",0.7],["control",0.7],["alter",0.4],["influence (halo)",0.2],["suppress",0.1],["convert",-0.0],["understand (control)",0.0],["describe (control)",0.0]]},{"id":"L19_E23_N321","k":"m","L":19,"E":23,"N":321,"t":"A list-slot neuron whose sign follows the topic","l":"This neuron reads the second item of a tight list or symbolic pair, the B in 'A, B' or 'A = B', and the topic decides its sign. Permission verbs ('read, update'), legal harm words, chord names and chosen indices push it positive. Natural-science lists, drug-name pieces and…","st":"tested","i":3,"c":4,"m":[["permission verbs",0.6],["indices",0.6],["chords",0.4],["CRUD variant",0.1],["halo materials",0.1],["harm adjectives",0.0],["control: food",-0.0],["control: colors",0.0],["geometry",-0.2],["science triad",-0.6]]},{"id":"L19_E254_N229","k":"m","L":19,"E":254,"N":229,"t":"A neuron that places a dead person's estate under a jurisdiction","l":"This neuron in Qwen3.6-35B-A3B reads succession law. It fires when text ties a dead person's estate to the place or law that governs it, as in 'where the decedent resided' or 'property located in', reaching about +11. Living holders with assets abroad, feudal lands held under…","st":"tested","i":6,"c":6,"m":[["late owner",4.7],["incapacity",4.5],["owner (past)",4.3],["dead testator",4.0],["decedent",3.5],["traveler",2.7],["living settlor",2.3],["hotel guest",2.1],["company",1.3],["divorce",0.9]]},{"id":"L19_E31_N303","k":"m","L":19,"E":31,"N":303,"t":"Who decides the election: a neuron for handing it to the House","l":"This neuron tracks who settles an election. It fires hardest on the word that hands an undecided election to a chamber: 'no candidate had a majority, so the election was thrown into the House.' The same effect appears in Spanish and German. It fires weakly when any body chooses…","st":"tested","i":7,"c":6,"m":[["no majority",7.9],["tie",7.3],["body chose",5.4],["clear winner",4.3],["split field",4.0],["contestant tally",2.0],["control: weather",1.8],["control: band",1.8],["control: kitchen",1.8]]},{"id":"L19_E42_N369","k":"m","L":19,"E":42,"N":369,"t":"A neuron that flags authors speaking through someone else's mouth","l":"This neuron in Qwen3.6-35B-A3B tracks whose voice a passage's words are in. It swings sharply negative, down to about -13, at the moment a text says an author is putting words into a speaker they created or borrowed: 'puts into the mouth of', 'Plato, through Socrates', 'Virgil…","st":"tested","i":9,"c":8,"m":[["own voice",0.2],["control: birth",0.0],["control: list",-0.0],["presents argument",-0.3],["attributes",-1.2],["speaks through",-1.3],["dialogue form",-3.2],["puts into",-5.0],["via characters",-7.8],["into the mouth",-8.6]]},{"id":"L19_E46_N304","k":"m","L":19,"E":46,"N":304,"t":"Stripes on a person, not silk in a list","l":"This neuron responds to what a described person looks like. It fires on patterns and markings carried by a person or god, such as stripes, spots, feathers or Hermes' wings, and stays flat on the garments themselves. It turns negative when a noun names a kind of stuff: fabrics…","st":"tested","i":4,"c":6,"m":[["stars",3.7],["stripes",2.8],["flowers",2.8],["spots",2.3],["feathers",2.2],["ribbons",1.2],["material",0.8],["control",0.0],["garment",-0.3],["accessory",-0.4]]},{"id":"L1_E11_N351","k":"m","L":1,"E":11,"N":351,"t":"A layer-1 neuron that spells out 'tempo', in any meaning","l":"In Qwen3.6's first layers, this neuron is a spelling detector for one word: 'tempo'. The token that completes 'tempo' drives it strongly negative, near −7, whatever the word means: musical pace, the Edo-era Tempo Reforms, a chemical oxidant or a startup. Close misspellings and…","st":"tested","i":4,"c":8,"m":[["p-os halo",0.7],["temp- control",-0.0],["synonym",-0.0],["synonym",-0.0],["temp- word",-0.9],["Italian plural",-1.2],["one-letter swap",-2.1],["p→b swap",-2.1],["plural",-5.1],["tempo (core)",-7.7]]},{"id":"L1_E201_N34","k":"m","L":1,"E":201,"N":34,"t":"Member after a dot: lowercase receiver positive, Capitalized negative","l":"In code, this neuron fires on the member name that follows '.' or '::', and its sign depends on how the receiver is spelled. A plain lowercase receiver gives a positive value (registry.register, fmt::Display). A Capitalized or ALL-CAPS receiver gives a negative one, deepest for…","st":"tested","i":4,"c":8,"m":[["module.Class",2.1],["lowercase var",1.8],["module::Type",1.2],["compound receiver",1.1],["no receiver",0.0],["prose control",0.0],["lowercase + parse",-1.2],["Capitalized var",-1.6],["Type::parse",-3.4],["PHP facade",-4.1]]},{"id":"L1_E20_N378","k":"m","L":1,"E":20,"N":378,"t":"A colon detector for inline citation tags like [source:","l":"This layer-1 neuron fires on the colon that opens an inline reference tag after a claim, as in '[source:', '[url:', '[citation:', or Chinese '来源:'. Those colons are the only tokens that open its gate. Web addresses and glued sentence starts produce only small leaks of either…","st":"tested","i":4,"c":8,"m":[["credit",0.7],["citation",0.6],["url",0.6],["link (weak)",0.5],["warning (control)",0.2],["tip (control)",0.1],["note (control)",0.1],["example (control)",0.1],["source",0.0],["via",-0.0]]},{"id":"L1_E213_N263","k":"m","L":1,"E":213,"N":263,"t":"A 漏 / draw / leak detector that turns negative on 'great'","l":"This neuron mostly reads which token it is looking at. Its strongest trigger is the Chinese character 漏 (leak, water clock, omission), which can reach about +10. It responds more weakly to related 'seep through / reveal' words such as 透露出, and in English to draw used as a verb…","st":"tested","i":3,"c":8,"m":[["透露出",1.2],["verb draw",0.5],["漏 (core)",0.0],["leak",0.0],["reveal (halo)",0.0],["control verb",-0.0],["control verb",-0.0],["synonym pull",-0.1],["great (negative)",-0.5]]},{"id":"L1_E218_N350","k":"m","L":1,"E":218,"N":350,"t":"A CT/CX detector that only wakes up inside gibberish","l":"This layer-1 neuron fires on capital C-led two-letter tokens, best CT and CX, and goes negative on closing brackets such as ')))'. Its expert is routed only after a long stretch of gibberish, base64 or abbreviation runs, so every strong firing sits in junk text. In English the…","st":"tested","i":3,"c":7,"m":[["CX (core)",2.6],["CT (core)",2.4],["CL",1.4],["CS",1.1],["non-C KT",0.6],["lowercase ct",0.0],["DT control",0.0],["real word",0.0],["bracket ]",0.0],["stacked )))",-0.8]]},{"id":"L1_E230_N120","k":"m","L":1,"E":230,"N":120,"t":"A 'Related Topics' detector whose negative pole is DefaultEnum","l":"This layer-1 neuron fires on the head word of a short two-word label, and it fires hardest on 'Topics' in 'Related Topics' navigation footers (about +2.6). The preceding modifier decides whether the expert is routed and also boosts the neuron's value. The head word sets the…","st":"tested","i":3,"c":8,"m":[["Related Topics",2.9],["Similar Topics",2.6],["Popular Topics",2.0],["Related Posts",0.3],["control: Articles",0.1],["Main topics",0.0],["bare Topics",0.0],["control: Comments",-0.0],["CustomEnum",-1.0],["DefaultEnum",-1.3]]},{"id":"L1_E239_N416","k":"m","L":1,"E":239,"N":416,"t":"A neuron that marks finished addresses: 老师好 up, 请 down","l":"This layer-1 neuron marks the moment a Chinese speaker finishes naming a person they are addressing. It fires on the function token right after a person noun: 好 in 老师好, or the comma or colon closing a vocative or speaker label. It goes negative when a courtesy or request word…","st":"tested","i":4,"c":6,"m":[["greeting 好",1.7],["vocative comma",0.8],["请 after comma",0.8],["polite greeting",0.7],["question mark",0.1],["colon",0.0],["list comma",0.0],["control: 的",0.0],["control: 们",0.0],["thanks after comma",-0.8]]},{"id":"L1_E23_N200","k":"m","L":1,"E":23,"N":200,"t":"A neuron for split 地, strongest after 席, plus code's Local","l":"This early-layer neuron responds to the Chinese character 地 when the tokenizer leaves it as a separate token. The character just before it sets the strength: after 席 (as in 席地而坐) it fires several times harder than after most other characters. A weaker second trigger is the code…","st":"tested","i":3,"c":5,"m":[["席地 (core)",5.7],["other CJK",1.3],["土 halo",1.0],["天 variant",0.8],["席子 control",0.4],["place control",0.2],["席上 control",0.2],["俱 negative",-1.1],["梨 negative",-1.5]]},{"id":"L1_E242_N156","k":"m","L":1,"E":242,"N":156,"t":"A one-token neuron for Korean 참, in every sense","l":"This layer-1 neuron is essentially a detector for one Korean token: the stand-alone syllable 참. It fires strongly whatever 참 means (very, true, endure, sesame) and whatever language surrounds it. Words that merely begin with 참 are weaker. A faint halo of unrelated…","st":"tested","i":3,"c":10,"m":[["참 (core)",6.3],["참 no space",5.2],["merged 참고",1.8],["정말 (same meaning)",0.2],["Chinese 参考",-0.0],["Notice",0.0],["Then (control)",-0.0],["Add (control)",0.0],["Next (negative)",-0.0]]},{"id":"L1_E34_N153","k":"m","L":1,"E":34,"N":153,"t":"A seam detector for split 'Taal' spellings, reversed by Vietnamese 'ả'","l":"This early-layer neuron watches the seam between two word pieces. It fires when an 'a'-ending piece is followed by 'al', completing a split 'Taal'-type spelling; this mostly means the Philippine lake and volcano, Thaaliba, or Taalas. It goes negative when an o-ending piece is…","st":"tested","i":4,"c":8,"m":[["Ta|al core",6.7],["Th|aal",5.6],["caps core",5.4],["Arabic ta'al",2.4],["o-vowel near miss",2.2],["T control",0.1],["long prefix",0.0],["control",-0.0],["tho + ả",-4.3],["o + ả",-4.9]]},{"id":"L1_E38_N446","k":"m","L":1,"E":38,"N":446,"t":"A conflict-of-interest header detector gated by research-paper routing","l":"This layer-1 neuron detects the words that open a conflict-of-interest declaration. It fires on 'interest' in 'Conflicts of interest' or 'Competing interests', and on the 'losures' in 'Disclosures'. The 'losures' response is mostly spelling: it also covers Enclosures and…","st":"tested","i":3,"c":6,"m":[["Disclosures",5.3],["Conflicts of interest",3.5],["other interest",2.5],["financial",1.3],["Declaration of",0.7],["interest rate",0.5],["Competing interests",0.1],["control",0.1],["control",-0.0],["is often",-0.7]]},{"id":"L1_E67_N96","k":"m","L":1,"E":67,"N":96,"t":"A name-ending neuron: Scan|ia lights up, Sk|ij goes dark","l":"This early-layer neuron reads the second half of split-up names. It fires strongly when the ending is 'ania' or 'ia' after an s-cluster stem, so Scania tops everything, including SCANIA lettering on truck grilles in photos. It turns negative when the ending is 'ij', 'ig' or 'ik'…","st":"tested","i":3,"c":8,"m":[["Scania (core)",2.6],["Ascan|ia",2.6],["Sek|ania",2.0],["Fasc|ia",1.6],["Lyc|ia (halo)",-0.0],["whole-word -ia",0.0],["control brand",-0.0],["control place",-0.0],["Sh|ij",-1.1],["Sk|ij (negative)",-1.6]]},{"id":"L1_E78_N137","k":"m","L":1,"E":78,"N":137,"t":"Glued slashes push it up, Spanish 'desp' pulls it down","l":"This early-layer neuron has two faces. It fires positively on a slash glued to the end of a word, as in 'Igreja/' or 'compra/', mostly in non-English text. Names and Persian or Arabic words give the largest values. Glued hyphens and table pipes give weaker echoes. Its far…","st":"tested","i":3,"c":7,"m":[["name + slash",1.4],["place + slash",1.3],["glued hyphen",0.9],["glued pipe",0.9],["noun + slash",0.7],["digit slash",0.2],["spaced slash",-0.0],["glued comma",-0.0],["'desp'",-3.4]]},{"id":"L1_E89_N275","k":"m","L":1,"E":89,"N":275,"t":"A u-plus-dental spelling detector: Puth and Tamil புத்த","l":"This early-layer neuron checks spelling: it fires on a word piece that spells u followed by a dental consonant. Examples are the 'uth' in Charlie P|uth and the Tamil த்த in புத்தகம் (book), which reaches about +8. The router limits what we see, sending these pieces to the expert…","st":"tested","i":4,"c":8,"m":[["Tamil புத்த",7.6],["core Puth",6.5],["Puth- surname",6.5],["Pr|uth",5.8],["P|ith control",0.5],["P|eth control",0.3],["halo Pl|uto",0.1],["M|uth unrouted",0.0],["plain name",0.0],["Lao final",-1.3]]},{"id":"L1_E94_N190","k":"m","L":1,"E":94,"N":190,"t":"A signed lookup on compound tails, gated by the router","l":"This neuron is a signed dictionary for the second half of glued compounds. When the router sends it a tail like Processor, Safe or 花, the neuron outputs a positive value. Tails like OutputStream, Writer, 派 or a traceback's tilde underline come out negative. The router decides…","st":"tested","i":3,"c":5,"m":[["core variant",1.0],["brand tail",0.6],["core",0.4],["halo",0.1],["control",0.1],["control",-0.0],["control",-0.0],["mild negative",-0.2],["negative",-0.4],["negative core",-0.7]]},{"id":"L20_E106_N393","k":"m","L":20,"E":106,"N":393,"t":"Folk culture and Black liberation, and the Grove Press imprint","l":"This neuron carries two unrelated signals. Its positive side responds to text about a people's own culture and liberation, mainly folklore and African American freedom and migration history; for other peoples' folklore, the router usually never sends the token to this expert.…","st":"tested","i":4,"c":4,"m":[["peasant",2.5],["Haitian",2.0],["Black",1.9],["African American",1.7],["Polish",1.7],["Swedish",1.6],["Irish",1.4],["control: corporate",0.0],["control: finance",0.0]]},{"id":"L20_E12_N448","k":"m","L":20,"E":12,"N":448,"t":"Who am I to dare? A neuron for presumptuous initiative","l":"This neuron marks the moment someone takes it upon themselves to act or to claim a role, with words like dare, presume, undertake or set up as. It fires hardest when the text has first made the actor lowly: a poor shepherd who dares reads about twice as high as a general doing…","st":"tested","i":8,"c":6,"m":[["undertook",9.7],["set up for",9.4],["presumed attempt",7.9],["dared",6.5],["took upon",5.9],["began (halo)",2.3],["assigned role",1.2],["concupiscence",0.7],["control: state",0.3],["control: sleep",-0.2]]},{"id":"L20_E193_N166","k":"m","L":20,"E":193,"N":166,"t":"A neuron that asks whether capital letters matter for matching","l":"This neuron tracks one question: does capitalization matter when text is compared, matched or looked up? It fires on the verdict ('case-insensitive', Spanish 'sensible a mayúsculas', Chinese 区分大小写) and on the word just before it, even before 'case' is written. Case-folding code…","st":"tested","i":7,"c":10,"m":[["HTTP headers",11.6],["Linux files",5.4],["domain names",2.8],["SQL keywords",0.8],["emails",0.0],["time zones",0.0],["prices",0.0],["trees",0.0],["style rule",-0.3]]},{"id":"L20_E196_N510","k":"m","L":20,"E":196,"N":510,"t":"A neuron for rounding an operand up to an easy anchor","l":"This neuron marks a specific mental-math move: writing a round number in place of a slightly smaller operand already on the page, as in '47*59: 47*60'. It fires on the round number's final zero, and only when that substitution happens. The same '60' with no earlier 59 does…","st":"tested","i":10,"c":8,"m":[["58→60",2.3],["59→60 core",2.1],["55→60",2.1],["45→60 wide",1.2],["61→60 down",0.9],["65→60 down",0.7],["12→60 rescale",0.3],["30→60 double",0.0],["no substitution",-0.0]]},{"id":"L20_E213_N207","k":"m","L":20,"E":213,"N":207,"t":"A neuron that flags forbidden values and recoils from trig identities","l":"Deep in Qwen3.6's math reasoning, one neuron marks the moment a trace declares a particular value off-limits, such as 't = 1 is excluded' because a denominator vanishes or a shape degenerates. It also responds to 'as x approaches' a boundary. Its stronger, opposite pole is…","st":"tested","i":6,"c":6,"m":[["forbidden",2.5],["control: critical",2.3],["excluded",1.9],["not allowed",1.4],["asymptote (halo)",1.3],["invalid",1.0],["rejected",0.7],["control: integer",0.1],["trig identity",-2.2]]},{"id":"L20_E217_N356","k":"m","L":20,"E":217,"N":356,"t":"A phrase-closing neuron: positive in encyclopedias, negative in boilerplate","l":"This neuron fires on the word that completes a phrase, and the surrounding passage sets the sign. In encyclopedic writing about famous laws, works, thinkers and places, the completing word drives it positive: 'narrowly passed Congress' or a title-final 'Philosophy'. Inside…","st":"tested","i":5,"c":6,"m":[["classic work",0.0],["founding law",0.0],["natural history",-0.0],["contract topic",0.0],["middling civic",0.0],["business",0.0],["everyday doc",0.0],["NDA wording",0.0],["control",0.0],["control",0.0]]},{"id":"L20_E5_N301","k":"m","L":20,"E":5,"N":301,"t":"A cancer-concept reader with a faint cardiometabolic opposite","l":"This layer-20 neuron reads the concept of cancer in health text and pushes it hard negative. It fires on the English word, on translations such as Krebs and cáncer, and on the slots in 'Cancer is a group of…'. Related words like tumor give only weak dips. The positive side is…","st":"tested","i":5,"c":8,"m":[["cardiometabolic",1.0],["complication",0.5],["control",-0.0],["control",-0.0],["leukemia",-0.8],["German",-2.6],["tumor",-2.6],["cancer type",-3.9],["malignancy",-4.0],["cancer",-5.9]]},{"id":"L20_E71_N142","k":"m","L":20,"E":71,"N":142,"t":"A neuron for HPC batch-job requests and their directives","l":"This neuron tracks requests for HPC batch jobs: Slurm, PBS and similar scheduler scripts, code that writes or parses their directives, and prose about asking for nodes, memory or walltime. Within that text it fires on the openings of arguments, such as ' the', '#SBATCH' and…","st":"tested","i":6,"c":8,"m":[["PBS job",19.9],["SGE",13.4],["Slurm",11.8],["LSF",8.6],["Kubernetes",6.9],["HTCondor",4.6],["mpirun",1.6],["movie",0.1],["cron",-0.2],["Python",-0.2]]},{"id":"L20_E83_N476","k":"m","L":20,"E":83,"N":476,"t":"A boundary neuron that goes negative on sign-up cards and offers","l":"This neuron answers one question at text boundaries: is this text trying to sign the reader up? Its expert is routed mostly to newlines, punctuation and digits. At those tokens, blank subscription cards, membership offers, benefit bullets and mail-order prices drive it strongly…","st":"tested","i":5,"c":6,"m":[["control: record",-0.4],["event form",-0.5],["control: contact",-0.8],["no header",-2.5],["library card",-3.0],["lower-case card",-3.2],["membership",-4.1],["order offer",-4.5],["subscription card",-4.8]]},{"id":"L20_E87_N347","k":"m","L":20,"E":87,"N":347,"t":"Court judging a person versus paperwork owed to authorities","l":"This neuron tracks one question inside legal text: is a court about to pass judgment on a person, or does the person owe an administrative office some paperwork? It reads strongly positive at 'convicted by' just before 'the court' names the judge, and negative at duties such as…","st":"tested","i":5,"c":6,"m":[["convicted",0.0],["sentenced",0.0],["acquitted",0.0],["disciplined",0.0],["fined",0.0],["arrested (control)",0.0],["thanked (control)",0.0],["renew every",-0.0],["show permit",-0.0],["return permit",-0.0]]},{"id":"L21_E108_N299","k":"m","L":21,"E":108,"N":299,"t":"Spaceflight programs push it up, fixed built structures push it down","l":"This neuron carries a signed sense of what the text is about, and it shows up on small words, digits and commas rather than on topic words. It turns positive when the text recounts how a crewed spaceflight program proceeds, with missions flown, delayed, costed or cancelled…","st":"tested","i":6,"c":8,"m":[["Soyuz",3.0],["cloud seeding",2.4],["Apollo",2.2],["Shenzhou",2.2],["Gemini",1.9],["oil well",0.3],["bridge",-0.2],["LNG plant",-0.3],["control: cake",-0.4],["control: textbook",-0.5]]},{"id":"L21_E110_N145","k":"m","L":21,"E":110,"N":145,"t":"A neuron for writing handed out as punishment","l":"This neuron recognizes writing handed out as a punishment: lines written a hundred times, a court-ordered essay, an apology letter, a Chinese self-criticism. It needs two things together, a task that produces text and a context of wrongdoing. Either one alone gives only a weak…","st":"tested","i":7,"c":7,"m":[["penal essay",12.4],["read book",9.4],["write lines",8.3],["apology letter",7.7],["copy rules",6.0],["run laps",2.8],["cleanup",0.7],["control",0.4],["control",-0.4]]},{"id":"L21_E112_N154","k":"m","L":21,"E":112,"N":154,"t":"A per-unit rate detector for small numbers in word problems","l":"This neuron marks the moment a word problem binds an amount to a unit: the 'per', '/h' or 'every' right after a small number, as in '$40/h' or '8 letters per day'. It weakens as the number grows and ignores vague quantities. Weaker echoes appear where a budget share lands ('45%…","st":"tested","i":4,"c":6,"m":[["8 per",-0.0],["3 per",-0.0],["8 every",-0.0],["20 per",0.0],["800 per",-0.0],["1,000 per",-0.0],["8 a (control)",-0.0],["many (control)",-0.0],["no number",0.0]]},{"id":"L21_E148_N462","k":"m","L":21,"E":148,"N":462,"t":"A 'we' neuron that tells the speaker's side from the reader's community","l":"This neuron responds only to 'we', 'our' and 'us', and its sign depends on who 'we' refers to. When 'we' is a group speaking outward to readers, such as an essay-help site or a founder speaking for his company, it goes positive. When 'we' takes the reader in as a fellow parent…","st":"tested","i":9,"c":8,"m":[["quoted founder",6.5],["essay service",1.4],["neutral topic",0.6],["dictionary example",0.3],["'you' control",-0.0],["'they' control",-0.0],["'I' control",-0.1],["fellow runners",-1.0],["fellow parents",-1.6],["cared-for 'our'",-2.7]]},{"id":"L21_E191_N243","k":"m","L":21,"E":191,"N":243,"t":"A neuron for block bodies delivered in replies versus empty slots","l":"This neuron marks the moment a block of material is about to be handed over in a reply, such as a code fence under 'css:' or the line after 'Ingredients:' in a chat answer. The reply frame matters: identical code text gives about 2.7 under User/Assistant and about 0.3 in a…","st":"tested","i":5,"c":6,"m":[["css fence",2.9],["js fence",2.4],["json brace",0.6],["label newline",0.5],["prose break",0.1],["turn end",0.1],["prose label",0.0],["blank",0.0],["empty section",-1.2]]},{"id":"L21_E201_N55","k":"m","L":21,"E":201,"N":55,"t":"A signed lookup on memorized names: Huckleberry up, Oscar Wao down","l":"This neuron reacts to the later pieces of names and titles the model has memorized. The sign is a lookup, not a rule. 'Huckleberry' in Twain's novel reads strongly positive, while 'Wao' in Oscar Wao and 'Morrie' read strongly negative. The same word can flip: 'children' is…","st":"tested","i":3,"c":4,"m":[["Huck Finn",0.8],["Twain variant",0.5],["Bilbo (halo +)",-0.0],["halo −",-0.0],["fruit control",-0.0],["plain control",0.0],["Morrie",-0.4],["Rushdie title",-0.6],["Oscar Wao",-1.3]]},{"id":"L21_E229_N361","k":"m","L":21,"E":229,"N":361,"t":"A neuron for actual-versus-formal parameter binding","l":"This neuron tracks one idea in code explanations: the values a caller passes bind to the slots a function declares. It fires hardest on the word naming one side of that pairing, such as 'actual' or 'formal', and on 'parameter' itself, in many languages. It then pushes the model…","st":"tested","i":5,"c":7,"m":[["formal (core)",9.2],["actual (core)",8.2],["real",7.2],["supplied",5.5],["passed by (neg)",5.1],["count (control)",3.0],["input",2.9],["takes (neg)",2.9],["given",1.5],["same (control)",1.5]]},{"id":"L21_E30_N199","k":"m","L":21,"E":30,"N":199,"t":"A car neuron that separates selling a vehicle from guaranteeing it","l":"In text about cars, this neuron records how the car is being framed. It turns positive when a car is placed in a market, a model year or a class: 'sold in', 'model year 2004', an ALMS 'GT' entry. It turns sharply negative when the car is the object of a coverage guarantee, as in…","st":"tested","i":5,"c":6,"m":[["marketed",1.3],["sold",1.2],["built",0.8],["insured",0.4],["control: parked",0.4],["recalled",0.2],["leased",-1.3],["covered",-1.3],["warranted",-4.8],["guaranteed",-5.3]]},{"id":"L21_E31_N77","k":"m","L":21,"E":31,"N":77,"t":"A 'Wait, no' detector for self-correction in math reasoning","l":"This neuron marks the moment a reasoner takes something back. It fires on 'no' right after 'Wait,' or 'Hold on,' in a math reasoning trace, and it fires more the longer the reasoning before it. Plain negations, everyday self-corrections and labelled speakers barely move it. Its…","st":"tested","i":6,"c":4,"m":[["Hold on, no",0.5],["Wait, no",0.4],["bare No",0.1],["Wait, not",0.0],["Wait, nope",0.0],["Hmm, no",0.0],["Actually, no",0.0],["there is no",0.0],["Wait, yes",0.0],["cube notation",-0.0]]},{"id":"L21_E33_N432","k":"m","L":21,"E":33,"N":432,"t":"Dissidents and censored press versus satirists, cartoons and co-ops","l":"This neuron tracks speech and dissent under hostile states. It fires on human-rights-report sentences about the press, NGOs and political prisoners in countries like Eritrea or Iran, and on dissident writers who keep writing or get published abroad. Its opposite pole is comic…","st":"tested","i":5,"c":7,"m":[["prisoners",3.5],["radio",2.2],["banned press",2.1],["journalists",1.9],["hospitals",1.9],["satirists",1.8],["bakeries",1.8],["co-ops",1.7],["control",1.4],["NGOs",1.1]]},{"id":"L21_E34_N491","k":"m","L":21,"E":34,"N":491,"t":"Famous kingdoms taking roles versus British county offices","l":"This neuron tracks historical political geography with two signs. It turns positive on the small word right after a famous premodern state or ruler is given a role, capital or territory, and the better known the entity, the stronger the response. It turns sharply negative, about…","st":"tested","i":5,"c":6,"m":[["Golden Ring",2.2],["Rurikid Kiev",2.2],["Hanseatic",1.0],["Silk Road",0.2],["fictional realm",0.1],["marketing team",0.1],["software",-0.2],["Westmeath",-0.9],["county sheriff",-3.6],["knight of shire",-7.1]]},{"id":"L21_E35_N498","k":"m","L":21,"E":35,"N":498,"t":"A quote-attribution neuron the router only lets through near Greece","l":"This neuron marks the moment a finished quotation hands off to its source: the closing quote and the reporting verb that follows. That value only appears where the router sends tokens to its expert, and the router responds mostly to Greek and Balkan context. One word, such as…","st":"tested","i":7,"c":6,"m":[["question quote",2.0],["Greek quote → verb",1.9],["dash + wrote",1.4],["inscription",0.5],["bare 'said that'",-0.0],["Germany (unrouted)",0.0],["according to",-0.1],["no-quote control",-0.1],["classical citation",-0.3],["ancient ethnography",-4.8]]},{"id":"L21_E74_N208","k":"m","L":21,"E":74,"N":208,"t":"A neuron that anticipates the cement plug in a well","l":"This neuron anticipates plugs in wells. In oil and gas text about sealing or drilling out a well bore, it fires at the slot just before the plug, its length, or the object to drill out is named. It reaches about 10 there, and it does not need the word 'plug' itself. Removing…","st":"tested","i":6,"c":6,"m":[["plugged",8.7],["sealed",5.5],["capped",3.1],["cemented",1.9],["jettisoned",1.4],["abandoned",1.3],["grouted (halo)",0.6],["control",-0.2],["control",-0.2],["peeled",-0.3]]},{"id":"L21_E75_N175","k":"m","L":21,"E":75,"N":175,"t":"A neuron that measures how public office is held","l":"This neuron reads one thing: how a public office is held. When text describes whether a governing body's members are paid, part-time, how often they meet, or how they are elected, it swings strongly negative, deepest for state legislatures (down to about -18). The same words…","st":"tested","i":6,"c":5,"m":[["nurses (control)",-1.5],["firefighters",-2.8],["cashiers (control)",-3.3],["church synod",-3.5],["corporate board",-4.3],["school board",-5.7],["county commission",-6.2],["city council",-7.3],["legislators",-13.9],["state senators",-14.4]]},{"id":"L21_E88_N177","k":"m","L":21,"E":88,"N":177,"t":"A territory-or-product switch for proper names, seen in French text","l":"This neuron answers one question about a proper name: is it a territory or a product? Country, province and invented-kingdom names push it positive, and so do the articles and commas that open the next item in a list of them. Software, browser and AI-assistant names push it just…","st":"tested","i":6,"c":7,"m":[["country",5.4],["country",4.9],["province",3.2],["invented realm",1.7],["common noun",-0.2],["city",-0.4],["person",-1.2],["microstate",-1.6],["AI tool",-4.4],["browser",-7.2]]},{"id":"L21_E89_N78","k":"m","L":21,"E":89,"N":78,"t":"A lone attacker's hands versus a master's servants doing the grabbing","l":"This neuron tracks whose hands are on a person in a story. It turns positive when one attacker grabs, drags or strangles someone, peaking on the 'by' in 'grabbed him by the collar.' It turns much more strongly negative when the hands belong to someone's crew of guards, servants…","st":"tested","i":7,"c":6,"m":[["father",2.7],["lone woman",1.9],["stranger",1.7],["rebels (group)",1.1],["officer",0.7],["police (group)",0.2],["friends",-1.1],["guards",-1.4],["bodyguards",-1.7],["servants",-1.7]]},{"id":"L22_E104_N467","k":"m","L":22,"E":104,"N":467,"t":"A neuron for cued completions, and one negative footer word","l":"This neuron completes learned pairings. It fires positively on a specific word only when specific earlier words call for it, most clearly 'represent' after 'warrant, guarantee,'. Change either the word or the cue and the firing disappears. Its strongest and most consequential…","st":"tested","i":3,"c":4,"m":[["represent (core)",2.1],["endorse",0.4],["promise",0.4],["verify",0.4],["certify",0.2],["weekday",0.0],["colour",0.0],["function word",0.0],["source (no footer)",-0.2]]},{"id":"L22_E108_N431","k":"m","L":22,"E":108,"N":431,"t":"Where things sit versus what a structure holds","l":"This neuron tells apart two ways of describing a structure. When text says where items sit (digits in the hundreds place, parity bits at numbered positions, cells on a Turing tape, memory addresses), it goes strongly negative, down to about −7. When text says what the structure…","st":"tested","i":6,"c":6,"m":[["set (core +)",1.7],["collection",0.7],["store",0.3],["capacity",0.1],["control",-0.0],["control",-0.0],["tape squares",-0.1],["bit positions",-1.2],["slots",-1.4],["memory cells",-2.0]]},{"id":"L22_E114_N128","k":"m","L":22,"E":114,"N":128,"t":"A neuron for 'I don't know what…'","l":"This neuron marks the moment someone says they don't know. It fires on the word right after 'I don't know', 'didn't know' or 'not sure', usually 'what', and fires strongest in speechless dialogue like 'I don't know what to say'. The same words in affirmative or rhetorical form…","st":"tested","i":5,"c":8,"m":[["don't know",2.9],["not sure",2.9],["unsure (halo)",2.5],["no idea",2.0],["didn't know",1.0],["sure (trace)",0.1],["do know",0.1],["wonder",0.0],["didn't care",-0.1]]},{"id":"L22_E120_N266","k":"m","L":22,"E":120,"N":266,"t":"A negative marker for line starts in framed formal documents","l":"This neuron goes negative at line starts inside formal documents: the break after a chapter title, the wrapped second line of an all-caps bill heading, or a blank date field in a legal form. The router decides which layout positions get a value at all, and the document frame…","st":"tested","i":3,"c":6,"m":[["preface",-0.0],["caps section",-0.0],["plain chapter",-0.0],["list row",-0.0],["formula",-0.0],["prose control 2",-0.1],["prose control",-0.3],["chapter heading",-1.1],["chapter I",-1.2]]},{"id":"L22_E125_N140","k":"m","L":22,"E":125,"N":140,"t":"A boundary neuron: code blocks positive, memorized boilerplate negative","l":"This neuron reads boundary tokens (newlines, indentation, periods) and gives them a sign. It fires most where a new block of simple Python code begins, and more weakly at line ends in student notes and verse. It turns negative inside text the model recognizes verbatim as a…","st":"tested","i":5,"c":6,"m":[["def header",1.2],["nested indent",0.7],["verse line",0.7],["math",0.6],["class header",0.5],["after imports",0.5],["slide title",0.3],["plain prose",0.1],["IUCN citation",-0.0],["GFDL footer",-1.9]]},{"id":"L22_E12_N391","k":"m","L":22,"E":12,"N":391,"t":"Posited by the writer or met by a character: one sign","l":"This neuron sits on indefinite noun phrases and signs them by who owns them. When the writer puts a phrase forward as a step in its own reasoning ('a theory posits a', 'we assume a', 'it's easy to think of an'), the article reads positive. When the phrase is something a…","st":"tested","i":6,"c":8,"m":[["proposed",0.5],["hedged",0.3],["posited",0.2],["control",-0.0],["control",0.0],["think of",-0.3],["assumed",-0.5],["found",-0.7],["saw",-1.3],["expected",-1.6]]},{"id":"L22_E140_N409","k":"m","L":22,"E":140,"N":409,"t":"A neuron for weapons being armed or hidden","l":"This neuron fires when a text puts a weapon, trap or harmful device together with an act that arms or hides it: loaded, rigged, planted, concealed. It fires on whichever of the two comes second. A bomb that is merely found or defused barely moves it. Harmless hiding earns a…","st":"tested","i":5,"c":6,"m":[["rigged",4.2],["concealed",2.7],["hidden",2.3],["disguised",2.1],["planted",0.7],["stored (halo)",0.0],["found (control)",-0.0],["removed (control)",0.0],["defused (control)",-0.2],["opened",-1.2]]},{"id":"L22_E164_N371","k":"m","L":22,"E":164,"N":371,"t":"A neuron for federations of unions, sports bodies, and churches","l":"This neuron tracks a sentence whose subject is a peak federation, a body whose members are themselves organizations, named on familiar templates like 'Trades Union Congress', 'National Olympic Committee' or 'Council of Churches'. It fires most on what is said about that body…","st":"tested","i":6,"c":8,"m":[["church council",7.7],["TUC (core)",7.3],["ICTU (core)",5.7],["fictional TUC",4.7],["Olympic body",4.3],["employer fed.",3.2],["single union",1.3],["learned society",0.6],["gov. agency",-0.3],["trade assoc.",-0.5]]},{"id":"L22_E181_N381","k":"m","L":22,"E":181,"N":381,"t":"A neuron that marks where paths begin, not where things sit","l":"This neuron tracks the role a place plays in a sentence, not the place itself. It fires on the head of a phrase naming where a path starts: a bus line's first stop, a ship's departure port, the landmark a distance is measured from, or one end of a line drawn across a map or…","st":"tested","i":7,"c":6,"m":[["distance from",4.8],["route origin",4.2],["corridor",4.1],["destination",2.9],["control: no place",0.1],["control: topic",-0.3],["static beside",-0.7],["sky position",-2.0]]},{"id":"L22_E224_N219","k":"m","L":22,"E":224,"N":219,"t":"The field word after 'University Institute for…'","l":"This neuron marks the field word in an institutional unit's name when 'university' modifies the unit, as in 'University Institute for Advanced Study' or 'Instituto Universitario de Investigación'. It responds weakly to other 'Institute for X' names and ignores 'University of X'.…","st":"tested","i":5,"c":6,"m":[["Univ Institute",5.8],["Univ Corporation",5.0],["lower-case univ",3.7],["no 'university'",1.5],["Max Planck",1.1],["Harvard name",0.6],["Univ business",0.2],["National",-0.0],["control",0.0],["book series",-0.9]]},{"id":"L22_E237_N494","k":"m","L":22,"E":237,"N":494,"t":"A neuron that anticipates the named cause of physical processes","l":"This neuron anticipates the cause of a physical process. In science prose that has just described something changing, it fires on the small word that opens the explanation, such as 'arise via the', before the mechanism's name appears. Context sets its sign. The same words become…","st":"tested","i":5,"c":6,"m":[["stellar pulse",4.6],["binary period",2.9],["granular halo",1.3],["ferry control",0.1],["prices control",-0.0],["poetry control",-0.0],["magnetic field",-0.2],["precision list",-0.7],["perturbative",-0.8]]},{"id":"L22_E252_N290","k":"m","L":22,"E":252,"N":290,"t":"What fills a fronted container, as opposed to an open vista","l":"This neuron fires when a sentence first puts a container or source up front ('the spheres from which…', 'in which…', 'Within this stone…') and then names what fills it, leaves it, or lives in it: air pumped out, a nucleus removed, ghosts, or even nobody. Swap the container for…","st":"tested","i":5,"c":4,"m":[["nucleus removed",4.5],["air removed",3.5],["ghost dwellers",1.8],["empty inside",1.2],["indwelling spirit",1.1],["ordinary residents",0.4],["plain order",0.0],["control",0.0],["instruction",-0.2],["open vista",-1.2]]},{"id":"L22_E255_N428","k":"m","L":22,"E":255,"N":428,"t":"A neuron for defined topics returning after 'of' or 'in'","l":"This neuron fires when a topic that an encyclopedia-style passage has already defined comes back as the domain of a phrase, as in 'Argentina is a country… The geography of Argentina'. A defining sentence roughly quadruples the response, from about +0.7 to +2.75. Cited-title…","st":"tested","i":4,"c":6,"m":[["defined topic",2.8],["related region",-0.0],["undefined country",0.0],["other continent",0.0],["learned term",0.0],["scholarly name",0.0],["opaque ID",-0.0],["vague noun",-0.0],["function word",0.0],["pronoun",0.0]]},{"id":"L22_E37_N469","k":"m","L":22,"E":37,"N":469,"t":"A neuron for template slots and memorized strings, signed item by item","l":"This neuron responds to slots in familiar technical text. It fires on the unknown coefficient letter in a factoring template, whatever the letter. It also fires on pieces of real, recalled strings such as paper titles and toll-free numbers, where each string sets its own sign…","st":"tested","i":3,"c":2,"m":[["real title",0.7],["real variant",0.3],["close variant",-0.0],["real, other",0.0],["real, negative",-0.0],["invented",-0.0],["invented",-0.0],["famous quote",0.0],["plain phrase",0.0]]},{"id":"L22_E4_N434","k":"m","L":22,"E":4,"N":434,"t":"A signed reader of identifier pieces: descriptive heads versus scope flags","l":"This neuron reads pieces of identifiers in code and config, and its sign says what kind of piece each one is. The head word of a familiar descriptive name, such as Timer in isIdleTimerDisabled or the restoration part of face_restoration, pushes it positive. Scope and placement…","st":"tested","i":3,"c":6,"m":[["face_restoration",1.6],["synonym head",1.4],["weaker synonym",0.8],["other head",0.7],["control",0.1],["control",-0.3],["flag piece",-0.7],["scope flag",-1.3],["scope _global",-2.6]]},{"id":"L22_E52_N429","k":"m","L":22,"E":52,"N":429,"t":"A legal-standing dial: under the court's power versus owing duties","l":"In legal text, this neuron tracks which way legal power points at the party being discussed. It fires positive at the word where a party submits to a court's jurisdiction or is shielded from it, as in 'irrevocably submits to'. It fires negative, and harder, where a party is set…","st":"tested","i":5,"c":7,"m":[["submits to",1.4],["consents",1.4],["amenable to",1.3],["immune from",1.1],["exempt from",-0.0],["liable for",0.0],["control: owns",0.0],["control: sells",-0.1],["comply with",-0.9],["meet its",-0.9]]},{"id":"L22_E67_N358","k":"m","L":22,"E":67,"N":358,"t":"Element symbols or word initials: a chemistry-slot expectation neuron","l":"Inside chemistry text, this neuron predicts how the next short capital-letter string should be read. It goes positive when the context promises element symbols, as in 'water vapor (H2O)'. It goes negative when the letters will be initials or code letters, as in drug regimens or…","st":"tested","i":9,"c":9,"m":[["water vapor",2.6],["acid formula",1.7],["HCl gloss",1.2],["trophic label",0.9],["gas formula",0.6],["acronym slot",0.2],["name initial",0.0],["institution",-0.0],["drug initials",-1.8],["cement code",-6.5]]},{"id":"L22_E81_N352","k":"m","L":22,"E":81,"N":352,"t":"Speaking and hearing as abilities, not audio devices","l":"This neuron tracks a person's own ability to hear, speak or express something openly. It fires on the small connecting words ('does', 'or', 'never') in such clauses, and fires most when earlier sentences have set up a limit such as deafness or privacy. Words like 'publicly' and…","st":"tested","i":4,"c":5,"m":[["overtness",1.5],["radio device",1.2],["explicit",0.7],["hear (core)",0.3],["speak",0.2],["audio device",0.1],["ability modal",0.0],["other sense",-0.0],["control",0.0],["control",-0.0]]},{"id":"L23_E130_N197","k":"m","L":23,"E":130,"N":197,"t":"A signed dial for what things are made of","l":"This neuron fires on the word that opens a slot saying what something has, is made from, or must satisfy, and the framing word sets the sign. A plain connector inside a requirement or making frame turns it positive, as in 'cannot be made by', 'without using', or 'if it has'.…","st":"tested","i":4,"c":6,"m":[["cannot be obtained by",4.3],["cannot be made by",4.1],["can be made by",2.9],["have",1.1],["come with",0.1],["control: bright",0.0],["control: look on",0.0],["formed from",-1.5],["consist",-1.9],["composed",-3.0]]},{"id":"L23_E144_N371","k":"m","L":23,"E":144,"N":371,"t":"Checkpoint filenames up, memorized web boilerplate down","l":"This neuron responds to strings the model has memorized. It fires most on saved machine-learning checkpoint names such as 'model.ckpt' and more weakly on other memorized code paths and prose templates. Other memorized boilerplate, such as texashistory.unt.edu citations…","st":"tested","i":3,"c":5,"m":[["Keras .h5",2.1],["model.ckpt",2.0],["other .ckpt",1.9],["pickle .pkl",1.1],["torch .pth",0.3],["control .csv",0.1],["memorized string",0.0],["control prose",0.0],["control .txt",-0.1],["version string",-1.0]]},{"id":"L23_E156_N270","k":"m","L":23,"E":156,"N":270,"t":"A neuron for pro-forms that close a phrase from memory","l":"This neuron fires when a small word such as 'one', 'none', 'any' or 'just that' ends its phrase by standing in for a kind named earlier, as in 'wanted a clock ... bought one'. The same words go negative when their content is still to come, as in 'enter to win one [of the ...]'…","st":"tested","i":6,"c":6,"m":[["count",0.6],["quantity halo",0.3],["kind 'one'",0.0],["'of one' frame",0.0],["pronoun control",0.0],["repeated noun",0.0],["no object",0.0],["modified 'one'",-0.1],["open 'the one'",-0.3]]},{"id":"L23_E156_N373","k":"m","L":23,"E":156,"N":373,"t":"Far and to-the-whole go up; along-the-entire and most-of go down","l":"This neuron responds to scope words such as far, whole, entire, most and much, and gives them a sign according to their role in the sentence rather than the word itself. It is positive when the word gives reach: how far away a place is, or everyone an action is addressed to. It…","st":"tested","i":5,"c":6,"m":[["glide far",2.3],["river far",1.1],["an immense",0.4],["to the whole",0.3],["an incredibly",0.1],["explored most",-0.0],["control: a",0.0],["control: the",0.0],["spent much",-0.3],["along the entire",-2.3]]},{"id":"L23_E173_N298","k":"m","L":23,"E":173,"N":298,"t":"A neuron that marks newly introduced topics, opposed to fixed names","l":"This neuron marks the moment a newly introduced topic is complete and the text is about to say something about it. It fires at the start of a new word problem, encyclopedia entry, book title or log line. It falls silent if the same wording appeared just before. Its negative side…","st":"tested","i":4,"c":6,"m":[["subject + has",1.3],["new problem",0.8],["code request",0.5],["title head",0.4],["control",-0.0],["control",0.0],["times-as chain",-0.9],["mid-number digit",-0.9],["species compound",-1.4],["colour modifier",-1.5]]},{"id":"L23_E178_N231","k":"m","L":23,"E":178,"N":231,"t":"A neuron that flags risky statements inside try and retry loops","l":"This neuron marks the end of a statement that might fail and is being guarded: a conversion, input read, file open or import inside a Python try body, or a cin read inside a C++ do-while retry loop. A literal assignment in the same spot does not fire. Straight-line statements…","st":"tested","i":6,"c":6,"m":[["math.log",3.6],["open file",3.3],["int(input())",2.5],["plain call",2.5],["float()",2.2],["indexing",0.9],["string literal",0.1],["literal 0",-0.1],["flag False",-0.1]]},{"id":"L23_E205_N222","k":"m","L":23,"E":205,"N":222,"t":"A neuron for the next term of a problem setup","l":"This neuron fires inside structured setups such as word problems, task instructions and reference lists, on the token that opens the next term of a schema already laid out. The token chosen matters as much as the slot: in a flow-rate question, 'runs' drives it about twenty times…","st":"tested","i":2,"c":1,"m":[["runs (core)",1.5],["flows",0.4],["passes",0.2],["enters",0.0],["goes",0.0],["falls (likely)",0.0],["will (control)",0.0],["is (control)",0.0],["they (control)",0.0]]},{"id":"L23_E213_N104","k":"m","L":23,"E":213,"N":104,"t":"Expected words in memorized slots, each with a fixed sign","l":"This neuron responds to a short list of particular words, but only when each word lands in the slot that the text before it has set up. 'Numbers' after 'Given a list of' and 'in' in an encyclopedia's public-domain footer push it up. 'British' in 'Handbook of British Chronology'…","st":"tested","i":3,"c":5,"m":[["numbers",0.4],["row of keys",0.3],["integers",0.1],["statement",0.1],["strings (weak)",-0.0],["control",0.0],["control",-0.0],["question after list",-1.0],["question slot",-1.4]]},{"id":"L23_E226_N142","k":"m","L":23,"E":226,"N":142,"t":"A neuron for nouns that fit their story","l":"This neuron in Qwen3.6-35B-A3B fires on the noun that closes a phrase when that noun suits its context: 'horse' in a horseman tale, 'train' in a train problem, 'Economy' in a scholarly book title. The same noun in the wrong context does nothing. Its negative side is a short list…","st":"tested","i":4,"c":2,"m":[["core title",1.7],["field noun",1.0],["field noun 2",0.9],["modifier halo",0.5],["off-frame noun",0.5],["task words",-0.0],["control: year",-0.0],["plain 'these'",-0.6],["'These' title",-0.8]]},{"id":"L23_E246_N443","k":"m","L":23,"E":246,"N":443,"t":"A title-noun closer that turns negative on abstract titles and names","l":"This neuron fires at the end of a head noun in a title, citation or dress-code phrase, and fires hardest on the subword piece that finishes a split word, such as Wrongful Convict|ions. The sign depends on what the title is about. Concrete or historical heads read positive, while…","st":"tested","i":4,"c":5,"m":[["attire",0.6],["split plural",0.0],["split plural",0.0],["split noun",0.0],["single token",0.0],["place name",0.0],["open modifier",-0.0],["control: function",0.0],["control: year",-0.0],["abstract title",-1.4]]},{"id":"L23_E60_N6","k":"m","L":23,"E":60,"N":6,"t":"A neuron for enterprise, OEM and private layers over shared software","l":"This neuron only computes on software text, because its expert is only chosen there. Inside that text it goes sharply negative on words that name a restricted or owned layer over a shared core: enterprise editions, OEM customizations, clients' private forks, local override…","st":"tested","i":5,"c":8,"m":[["premium",-0.0],["community (control)",0.0],["free (control)",-0.0],["open (control)",-0.7],["local layer",-1.6],["identifier halo",-2.1],["enterprise",-2.8],["private fork",-2.8],["EE",-3.1],["proprietary",-4.2]]},{"id":"L23_E66_N133","k":"m","L":23,"E":66,"N":133,"t":"A neuron for the verb that opens a verdict","l":"This neuron marks the moment someone begins handing down a judgment. It fires on believe, feel or think (and their equivalents in Chinese, Spanish and Arabic) when the context has asked for an evaluation, such as feedback letters, reviews, 'anything you feel should be added', or…","st":"tested","i":6,"c":8,"m":[["believe",7.8],["feel",5.6],["think",4.4],["consider",4.3],["suppose",3.7],["know (control)",0.2],["ate (control)",-0.0],["noticed (control)",-0.3],["hope (control)",-0.4]]},{"id":"L23_E71_N441","k":"m","L":23,"E":71,"N":441,"t":"A weak neuron: memorized strings versus named-source slots","l":"This neuron is weak and mixed, and switching it off barely changes the model's output. Its sign tracks two kinds of familiar slot. It goes positive when a token continues a string the model has memorized, such as a popular hex colour code or one specific Montessori lesson text.…","st":"tested","i":3,"c":3,"m":[["hex prefix",1.5],["building",0.4],["newspaper",0.2],["plain noun",-0.0],["BERT",-0.0],["AI model",-0.0],["name in list",-0.6],["music app",-0.7],["learning app",-1.2]]},{"id":"L23_E73_N356","k":"m","L":23,"E":73,"N":356,"t":"A neuron that marks remembered passages, signed by their subject","l":"This neuron fires late in passages the model has memorized, at the word where the familiar wording locks in. Its sign depends on what the passage is about. Physics history, health-research credits, contest stories and docstrings push it positive. Wikipedia definitional leads…","st":"tested","i":4,"c":6,"m":[["Einstein (core)",1.5],["baker (ctrl)",0.1],["Darwin (halo)",-0.0],["Augustine (neg)",-0.0],["Luther (neg)",0.0],["Ford (assembly)",-0.0],["neighbor (ctrl)",-0.0],["Bohr",-0.1],["Planck",-0.2]]},{"id":"L23_E96_N189","k":"m","L":23,"E":96,"N":189,"t":"A '4' ending numbers in code and data dumps","l":"This neuron watches for the digit '4' at the end of a number written as data: array elements, hashes, IDs and catalog codes. On that token it pushes strongly negative, most of all when the number is '734'. The same number in ordinary prose barely moves it. It also has a weak…","st":"tested","i":3,"c":2,"m":[["word control",0.0],["zero control",-0.0],["134",-0.2],["731 (not 4)",-0.3],["234",-0.5],["634 (weak)",-0.5],["738 (not 4)",-0.8],["784",-3.0],["734 (core)",-4.4]]},{"id":"L23_E98_N231","k":"m","L":23,"E":98,"N":231,"t":"A neuron that separates thoughtful advice from routine commands","l":"This neuron tracks instructions addressed to the reader, and its sign tells two kinds apart. It goes positive when a stated goal turns into advice the reader must think through, as in 'To assess whether you owe more, consider…'. It goes negative on routine commands that are…","st":"tested","i":5,"c":7,"m":[["list step",0.6],["epistemic goal",-0.0],["epistemic goal 2",-0.0],["practical goal",-0.0],["bare connective",-0.0],["modal control",-0.0],["first-person control",-0.0],["past-tense control",-0.0],["lesson direction",-1.6],["rules list",-2.5]]},{"id":"L24_E100_N343","k":"m","L":24,"E":100,"N":343,"t":"A slot reader that fires on stock templates, with either sign","l":"This neuron reads the slot of a familiar, heavily repeated template, and the template decides the sign. Words in cited titles (the 'A Dialogue' of a book subtitle), digits in stock math problems, and memorized code constants push it positive. A short list of memorized lists and…","st":"tested","i":3,"c":4,"m":[["stock subtitle",2.5],["later title word",2.5],["punctuation",2.5],["synonym",1.4],["generic head",0.8],["generic head",0.6],["near synonym",0.3],["function word",0.0],["article control",-0.0]]},{"id":"L24_E118_N440","k":"m","L":24,"E":118,"N":440,"t":"A Chinese neuron that waits for the head of 'possess the ability to…'","l":"This neuron tracks one Chinese construction: a verb of capacity or condition, such as 具备 (possess) or 养成 (develop), whose object is a modifier plus an abstract head like 能力 (ability) or 条件 (condition). It fires positive on the last word of the modifier, just before the head…","st":"tested","i":4,"c":8,"m":[["具备 + modifier",2.0],["养成 + modifier",1.5],["weak governor",0.4],["control: travel",-0.0],["control: plain verb",-0.1],["weakest governor",-0.2],["open 具有",-0.7],["purpose 达到",-0.7],["拥有 + modifier",-1.2],["concrete object",-1.2]]},{"id":"L24_E126_N371","k":"m","L":24,"E":126,"N":371,"t":"A seven-loving problem-number neuron that dislikes re-mentioned team names","l":"This neuron responds to the number that fills a quantity slot in a stock word problem, as in 'The average of seven…' or 'Rs 378', and it favors seven. It also fires on elements of number series. It turns negative when a proper name, a sports team most of all, comes back as the…","st":"tested","i":3,"c":5,"m":[["two-digit 7",1.7],["seven",1.6],["five",0.9],["nine",0.3],["two-digit 0",0.1],["common noun",0.0],["team name",0.0],["control",-0.0],["control",-0.0]]},{"id":"L24_E134_N136","k":"m","L":24,"E":134,"N":136,"t":"A neuron for recalling a name's spelling, not copying it","l":"This neuron marks the moment a model begins spelling a name or rare term it must pull from memory: a person, an institution, a quoted title, a chemical or species name. It is strongest when the name is rendered from a foreign-language source. It turns sharply negative when the…","st":"tested","i":6,"c":8,"m":[["work title",0.2],["person name",-0.0],["rare term",-0.0],["place name",-0.0],["other genus",-0.0],["control",-0.0],["control",-0.0],["bracket repeat",-0.0],["rank + name",-0.1],["plain repeat",-0.2]]},{"id":"L24_E144_N25","k":"m","L":24,"E":144,"N":25,"t":"A neuron for the upper limit of LaTeX sums","l":"This neuron reads the upper limit of a numeric range written in LaTeX math, such as the 3000 in a sum from k=1001 to 3000, and fires hardest on round competition-style bounds. Weaker echoes appear on repeated terms and on PPO training code. It dips on constraint clauses, Keras '…","st":"tested","i":4,"c":2,"m":[["sum to 3000",3.5],["sum to 2000",1.9],["repeated term",1.9],["arbitrary bound",1.8],["small range",0.8],["no lower limit",0.0],["symbolic bound",0.0],["prose range",-0.0],["work problem",-2.5]]},{"id":"L24_E186_N511","k":"m","L":24,"E":186,"N":511,"t":"A neuron for the hand-off moment: click, call, submit","l":"This neuron writes a 'submit / call / send' signal at the moment something is about to be sent into a system: an application into a ticket ballot, a click on a submit button, an email to a service desk, or a callback handed to the browser. It goes negative once the request has…","st":"tested","i":4,"c":5,"m":[["click",2.1],["submit",0.4],["request",0.2],["send",0.1],["call",0.1],["email",0.0],["share (halo)",0.0],["ask",0.0],["relax (control)",0.0],["smile (control)",-0.0]]},{"id":"L24_E196_N293","k":"m","L":24,"E":196,"N":293,"t":"A neuron that lights up when code ranks things by size","l":"Inside code and technical text, this neuron marks the moment a collection is being ranked by size. It reaches about +6 on idioms like argsort(x)[::-1], and it also fires on prose that says 'arranged in descending order'. The response builds up from three parts: a sort, a numeric…","st":"tested","i":5,"c":8,"m":[["value, desc",2.2],["rank by score",1.7],["by score",1.6],["by size",1.1],["by name",1.1],["by date",0.9],["no order",0.0],["count only",-0.2],["control",-0.2]]},{"id":"L24_E222_N391","k":"m","L":24,"E":222,"N":391,"t":"A 'used for' neuron that turns negative on bodily testing","l":"This neuron fires on the verb or answer slot that announces what something is for, such as 'lithium carbonate is used' or 'Why donate bone marrow? A:'. It fires most when the thing is a raw material headed into another product. It goes negative on text about bodies and their…","st":"tested","i":5,"c":5,"m":[["artemia",1.8],["krill meal",1.3],["graphite",1.3],["lithium",1.2],["hearing test",0.5],["copper ore",0.4],["tool (halo)",0.4],["cotton",0.3],["end product",0.2],["sea salt",0.1]]},{"id":"L24_E236_N490","k":"m","L":24,"E":236,"N":490,"t":"A neuron that flags tricyclic psychiatric drugs like chlorpromazine","l":"Inside one expert of a large mixture-of-experts model, this neuron mainly flags tricyclic psychiatric drugs. It fires strongly negative, near -15, on the middle of names like chlorpromazine, and that state carries onto the words that follow. A drug must be both tricyclic and…","st":"tested","i":7,"c":6,"m":[["control",0.1],["patch drug",-0.2],["tricyclic, non-psych",-0.6],["non-tricyclic psych",-2.3],["azine dye",-3.1],["phenothiazine",-4.4],["tricyclic AD",-4.4],["clozapine",-5.8],["imipramine",-6.0],["chlorpromazine",-11.4]]},{"id":"L24_E243_N303","k":"m","L":24,"E":243,"N":303,"t":"A nutrition neuron that marks the next share of a breakdown","l":"Inside nutrition and biochemistry text, this neuron marks the moment a composition breakdown moves on: the second, smaller share of a whole, or a new kind of item after a run of similar ones. The first item stays silent. Guideline nouns, plain food lists and polynomial…","st":"tested","i":4,"c":5,"m":[["small share",1.9],["second share",1.6],["further share",1.5],["large share",0.5],["new class",0.5],["food list",0.1],["repeat whole",-0.0],["bio halo",0.0],["guidelines",0.0],["off-topic",0.0]]},{"id":"L24_E255_N258","k":"m","L":24,"E":255,"N":258,"t":"From Neolithic 'to' Roman, against the Zealots","l":"This neuron is a signed axis inside an expert that handles ancient Mediterranean history. Its positive end marks the connective that spans ancient periods, the 'to' in 'from late Neolithic Age to Roman Greece'. Its stronger negative end marks the First Jewish-Roman War: Zealots…","st":"tested","i":6,"c":8,"m":[["Neolithic (core)",4.6],["Bronze Age",3.6],["Byzantine",3.5],["Ptolemaic",3.0],["Jomon (non-Med.)",1.5],["Hasmonean Judea",1.3],["non-period",-0.1],["modern year",-0.2],["Titus's siege",-2.0],["Zealot revolt",-2.3]]},{"id":"L24_E54_N436","k":"m","L":24,"E":54,"N":436,"t":"A neuron that notices headings stranded inside scanned history prose","l":"This neuron reacts to printed pages that have been scanned and flattened into text. It fires on the period or line break that closes a segment when a margin heading or caption has landed inside an unfinished sentence, especially when line-end hyphens split words nearby. It…","st":"tested","i":5,"c":6,"m":[["margin heading",4.2],["caption",4.0],["LaTeX markup",3.4],["short heading",3.3],["table header",3.3],["full sentence",3.0],["no insert",2.9],["no insert 2",2.9],["chapter title",2.8]]},{"id":"L25_E108_N177","k":"m","L":25,"E":108,"N":177,"t":"Pointing at unlettered answer options, and settling them","l":"This neuron tracks when a text refers to answer options by position rather than by letter, after a question has opened a choice. It peaks on quoted ordinal labels like \"first\" in 'Answer with \"first\" or \"second\"', on breaks between dash-bulleted options, and where a solution…","st":"tested","i":5,"c":6,"m":[["first",4.3],["fourth",3.8],["second",2.6],["Spanish ordinal",2.3],["numbered",0.9],["digit",0.7],["letter A",0.1],["yes/no",0.1],["true/false",0.0],["letter B",-0.2]]},{"id":"L25_E149_N392","k":"m","L":25,"E":149,"N":392,"t":"A neuron for 'twice as old as my [sister]'","l":"This neuron marks the person or thing a speaker measures themselves against in a ratio comparison. In 'I was twice as old as my sister' it fires on 'sister', and the router sends that token to this expert as its first choice. It turns negative when a money problem asks how much…","st":"tested","i":5,"c":5,"m":[["twice as rich",1.2],["twice as old",0.9],["half as old",0.9],["twice as tired",0.9],["twice as big",0.8],["older (no ratio)",0.8],["3x as old",0.5],["same age",0.5],["living with",0.0],["angry at",-0.0]]},{"id":"L25_E153_N128","k":"m","L":25,"E":153,"N":128,"t":"A neuron that separates a ruler's household from his state","l":"This neuron separates a pre-modern ruler's private household from his public state. It fires positively on continuations inside descriptions of eunuchs, harems, court musicians, imperial court music and staff counts. It fires negatively on governors, ministers, provinces and…","st":"tested","i":6,"c":6,"m":[["eunuchs",1.5],["harem",1.4],["chamberlains",1.2],["court music",1.1],["customs officials",0.9],["control: objects",0.6],["governors",-0.2],["control: plants",-0.3],["ministers",-0.8],["household staff",-1.9]]},{"id":"L25_E159_N309","k":"m","L":25,"E":159,"N":309,"t":"Who chooses among which products: a consumer-choice neuron","l":"This neuron marks the moment expository prose names who is choosing and what they choose among, as in 'individuals and groups' or 'products or services'. The neuron itself responds to choice frames in any domain. Its expert's router, however, admits mainly commerce and…","st":"tested","i":6,"c":6,"m":[["products or",3.1],["products/services",3.1],["tourism",1.0],["stores",0.4],["brands",0.2],["politics",0.2],["natural lang.",0.1],["plants",0.0],["control: math",0.0],["control: chem",0.0]]},{"id":"L25_E17_N101","k":"m","L":25,"E":17,"N":101,"t":"A neuron that goes dark for Taiyuan, Shanxi and Vojvodina","l":"This neuron is a negative detector for two specific regions. It plunges on mentions of Taiyuan and Shanxi, its institutions and the warlord Yan Xishan, reaching about -11. It also dips on Vojvodina in northern Serbia. Neighbouring Chinese provinces barely register. Its positive…","st":"tested","i":5,"c":7,"m":[["Hebei (ctrl)",0.1],["Xi'an (ctrl)",-0.0],["Manchester (ctrl)",0.0],["Hohhot (ctrl)",-0.2],["Datong",-0.5],["Changzhi",-0.6],["Novi Sad",-0.7],["Yangquan",-2.4],["Shanxi",-5.6],["Taiyuan",-7.2]]},{"id":"L25_E195_N395","k":"m","L":25,"E":195,"N":395,"t":"A neuron for things that won't fit, and things that won't cover","l":"This neuron tracks denied size relations and signs them by direction. It fires positively when something won't go inside a bound: a number that doesn't fit into a 32-bit integer, a trophy that won't fit into a suitcase, a topic beyond the scope. It goes negative when a holder…","st":"tested","i":7,"c":6,"m":[["doesn't fit into",6.0],["may not fit",5.4],["can't represent",4.5],["outside range",2.8],["beyond scope",2.4],["fits (control)",0.0],["stored (control)",0.0],["printed (control)",0.0],["not divisible",-0.7],["doesn't include",-0.9]]},{"id":"L25_E211_N454","k":"m","L":25,"E":211,"N":454,"t":"The earlier case's only member, named again in a parallel case","l":"This neuron fires when a later, parallel case names again the item that alone made up an earlier case, as in 'the cook feeds only the sailors… on cruise liners, meals for the crew'. It matches by meaning, not spelling, and is strongest when the second case widens the scope.…","st":"tested","i":6,"c":6,"m":[["partial match",8.7],["synonym (core)",8.4],["close variant",8.3],["close variant",7.1],["exact repeat",5.9],["pair baseline",3.3],["unrelated",1.5],["different group",0.4],["control",-0.0]]},{"id":"L25_E235_N327","k":"m","L":25,"E":235,"N":327,"t":"Caused consequence versus intended target: a sign-split causal connective neuron","l":"This neuron reads causal connectives, mostly in Chinese and Thai, and its sign tells what kind of cause is involved. When an event, condition or someone's fault brings a result about (你家开着窗，导致), it goes strongly negative, whether the result is good or bad. When a connective…","st":"tested","i":6,"c":6,"m":[["step 使",1.5],["step 到",0.7],["control",-0.0],["control",0.0],["promise 让",-0.1],["deliberate 让",-0.6],["event 让",-0.6],["mechanism",-1.9],["blame 导致",-4.4],["disaster",-4.9]]},{"id":"L25_E240_N35","k":"m","L":25,"E":240,"N":35,"t":"A neuron for class names like FRUITS that open sorted answers","l":"This neuron responds to a handful of literal kind nouns: fruits, viruses, planes, windows and the circum- of circumcenter. It responds most strongly when an assistant's answer opens by re-naming one class the user asked about, as in 'FRUITS'. Elsewhere its up term is negative…","st":"tested","i":3,"c":6,"m":[["fruits (core)",-0.0],["viruses",0.0],["planes",-0.0],["windows",-0.0],["vegetables (weak)",0.0],["trains (weak)",-0.0],["label (halo)",-0.0],["control",0.0],["control",-0.0],["digits (negative)",-0.0]]},{"id":"L25_E87_N195","k":"m","L":25,"E":87,"N":195,"t":"A neuron that marks both halves of a spelling-variant pair","l":"This neuron tracks pairs of spellings for the same word. When a word already in context is written again in its other spelling, it fires at the exact point where the spellings diverge: the 'oes' in oesophagitis after esophagitis. Made-up insertions do not count, and the router…","st":"tested","i":7,"c":6,"m":[["nonce insert",1.6],["UK variant",0.7],["close variant",0.7],["same spelling",-0.0],["gloss halo",-0.0],["unrelated word",0.0],["control",0.0],["nonce insert",-0.7],["next first form",-1.5]]},{"id":"L25_E97_N468","k":"m","L":25,"E":97,"N":468,"t":"A neuron that tells outcomes that happened from outcomes still awaited","l":"This neuron reads the stance of an outcome at the end of an action clause. On the result word, or the comma right after it, it turns positive when the outcome really happened ('the city went up in flames'). It turns negative when the outcome is only hoped for, searched for…","st":"tested","i":5,"c":6,"m":[["burned (core)",2.7],["burned",1.0],["destroyed",0.8],["non-violent result",0.6],["ongoing",-0.0],["threatened",-0.0],["seeking",0.0],["awaited",-0.0],["control",-0.0],["control",0.0]]},{"id":"L26_E111_N240","k":"m","L":26,"E":111,"N":240,"t":"A neuron for light-adapting lenses, opposed to lock-and-fitting hardware","l":"This neuron tracks the vocabulary of eye lenses that react to or block light: photochromic, polarized and self-tinting lenses, along with the goggles, masks and welding helmets that carry them. The words alone are enough; 'auto-darkening wood stain' still lights it up. Its…","st":"tested","i":4,"c":6,"m":[["auto-dark helmet",5.2],["photochromic",2.5],["bifocals",1.7],["photo dyes",1.7],["goggles",1.1],["polarized",0.9],["control: books",-0.0],["control: mugs",-0.1],["deadbolts",-0.7],["lantern parts",-1.6]]},{"id":"L26_E158_N417","k":"m","L":26,"E":158,"N":417,"t":"Pointing at marked pieces of text, versus adding to items","l":"This neuron reads whether a text or program is picking out pieces of a document that are already marked, or attaching something new to each item. It fires when an instruction points at annotated terms or a script selects highlighted code blocks, and it goes negative when code…","st":"tested","i":4,"c":6,"m":[["code blocks",4.9],["marked terms",3.7],["underlined words",3.3],["footnotes",3.1],["control",2.3],["images",1.8],["control",1.8],["write style",1.6],["buttons",0.7],["add field",-1.2]]},{"id":"L26_E167_N39","k":"m","L":26,"E":167,"N":39,"t":"Asking for help with hard family conversations, not being taught how","l":"This neuron responds when someone seeks guidance on a difficult conversation with another person, most strongly a painful family matter such as a parent's drinking or telling a child about death. It needs both a request for help and delicate interpersonal content. A frame alone…","st":"tested","i":6,"c":6,"m":[["asks suggestions",3.4],["asks advice",2.6],["instructions",2.1],["needs tips",1.6],["reported advice",1.1],["offered tips",0.6],["learn how",0.5],["taught me how",0.4],["control",0.3]]},{"id":"L26_E187_N177","k":"m","L":26,"E":187,"N":177,"t":"A neuron that anticipates the N+1 query problem","l":"This neuron tracks the N+1 query problem, where code loads related records with one database query per row instead of all at once. It fires hardest on the word just before the problem's name, as in 'includes(:comments) fixes the', and so helps the model predict 'N+1'. Swapping…","st":"tested","i":7,"c":8,"m":[["preload",17.9],["eager_load",17.8],["includes",16.6],["includes x2",16.6],["joins",13.3],["where",3.9],["select cols",2.4],["find_by",1.5],["order",0.5]]},{"id":"L26_E200_N189","k":"m","L":26,"E":200,"N":189,"t":"A slider-widget neuron with a few memorized side topics","l":"This neuron has no single abstract theme. It is a short list of memorized topic basins. The strongest is UI slider widget text: a component library's Slider, SliderTrack and SliderThumb, or Kendo tick options. Weaker basins are RabbitMQ plugin setup, writing-system reference…","st":"tested","i":3,"c":5,"m":[["Slider",2.4],["RangeSlider",2.3],["SliderThumb",1.9],["Progress",1.8],["NumberInput",1.5],["Welwitschia",1.2],["RabbitMQ",0.7],["Button (ctrl)",0.2],["Modal (ctrl)",0.1]]},{"id":"L26_E243_N256","k":"m","L":26,"E":243,"N":256,"t":"Limiting a person's exposure versus using a resource efficiently","l":"This neuron gives a signed reading at connectives such as ' of' and ' to' that follow a control measure. It turns positive when the measure limits what a person is exposed to or takes in: insulin doses, sleep restriction, noise limits, burn dressings. It turns more strongly…","st":"tested","i":4,"c":6,"m":[["burn first aid",1.2],["insulin dosing",0.0],["noise exposure",-0.0],["pest control",0.0],["control: read",0.0],["control: count",-0.0],["resource planning",-0.0],["wood efficiency",-0.0],["appropriate use",-1.1],["utilization rate",-1.3]]},{"id":"L26_E92_N153","k":"m","L":26,"E":92,"N":153,"t":"A neuron that spots the last value of a mirrored row","l":"This neuron watches for a list or table row that repeats an earlier one with shifted values. On the second row it fires on the digits of the endpoint cells, most strongly on the cell lined up with the first row's last item (Ben's 83 after Ana's 80). The same row with no earlier…","st":"tested","i":7,"c":6,"m":[["offset row",6.8],["muffin row",5.3],["no offset",4.4],["other offset",2.8],["partial row",1.7],["single value",0.5],["other domain",-0.4],["no earlier row",-0.4],["words only",-0.6]]},{"id":"L27_E113_N137","k":"m","L":27,"E":113,"N":137,"t":"A neuron that anticipates the verdict: was the feeling genuine?","l":"This neuron fires when a text is about to rule on whether someone's feeling, apology, belief or stated preference is genuine. It peaks on the word just before the verdict, and it peaks hardest after a staged display or doubt has raised the question. Doubt and negation drive it…","st":"tested","i":8,"c":8,"m":[["belief (core)",13.4],["grief",11.3],["apology",10.8],["gratitude",10.5],["stated preference",7.9],["remorse",7.3],["object control",4.7],["object control",2.6],["object (halo)",1.9],["neutral control",0.1]]},{"id":"L27_E151_N232","k":"m","L":27,"E":151,"N":232,"t":"A neuron for problems that slip past your own care","l":"This neuron fires in health, pet and home care advice at the moment the text is about to admit that a problem gets past your own care, for example 'Even with regular self-exams, some lumps can…'. It is strongest with that concession, around +6, and much weaker without it. It…","st":"tested","i":6,"c":7,"m":[["despite checkups",4.3],["despite brushing",3.7],["despite (variant)",3.5],["wrong-domain concession",3.2],["no matter how",2.9],["no care (halo)",2.7],["poor care (halo)",1.3],["acute onset",0.4],["unrelated",-0.2]]},{"id":"L27_E156_N232","k":"m","L":27,"E":156,"N":232,"t":"Radon kit placement and property-list slots, against preservation and square-unit qualifiers","l":"Inside text about land, buildings and rooms, this neuron marks the slot where the next item is about to be named. Its strongest case is the location answer after a radon test kit, as in 'place the radon kit [in]'. It also fires on the first separator in a list of property…","st":"tested","i":4,"c":6,"m":[["radon detector",3.9],["radon monitor",3.8],["radon kit (short)",3.4],["radon kit",3.2],["CO detector",-0.0],["smoke alarm",-0.0],["no radon",-0.0],["control",-0.0],["control",0.0]]},{"id":"L27_E177_N497","k":"m","L":27,"E":177,"N":497,"t":"Over parallel items: 'both' fires up, each-item results fire down","l":"This neuron only wakes up when the text has set up several parallel items, such as a length and a width, two polynomials, or a system of equations. Its sign then depends on the token's role. A word that spreads one operation over all the items, such as 'both' in 'cut from both…","st":"tested","i":5,"c":7,"m":[["to one",0.0],["single root",-0.0],["off-topic",0.0],["added to both",-0.1],["added to each",-0.1],["one of",-0.1],["control",-0.1],["between",-0.4],["other roots",-2.9],["third roots",-2.9]]},{"id":"L27_E17_N194","k":"m","L":27,"E":17,"N":194,"t":"One power covering a land versus features split between nations","l":"This neuron reads geographic and political text along one axis. It turns positive when a single power, such as an empire, a storm or a law, extends over a territory. It turns negative when a lake, river, pipeline or route sits between two or more countries or crosses from one…","st":"tested","i":6,"c":6,"m":[["rule reached across",3.2],["empire covered",2.9],["storm swept",2.6],["company-wide",2.0],["rains fall over",0.7],["located in",0.6],["control",0.1],["shared by",-0.5],["Russia into",-1.3],["Canada into",-2.6]]},{"id":"L27_E218_N362","k":"m","L":27,"E":218,"N":362,"t":"A neuron for where damaged text resumes after missing words","l":"This neuron marks the spot where damaged text picks up again: in scraped, spun or OCR pages that have already lost words, it fires on the token right after a fresh gap. Both conditions matter, since cleaning the earlier gaps nearly erases the response. Its opposite pole is the…","st":"tested","i":6,"c":7,"m":[["gapped caption",2.6],["math scratch",0.4],["filled caption",0.3],["line wrap",0.1],["clean prose",0.1],["C++ comment",0.1],["spun SEO",-0.3],["gapped history",-0.5],["man-page quote",-1.7]]},{"id":"L27_E233_N218","k":"m","L":27,"E":233,"N":218,"t":"A neuron for reaching after a name you can't quite recall","l":"This neuron fires when text reaches for a name it does not state: someone forgot it, asks to be reminded, or asks what a thing is called. Hesitant spoken groping ('uh, what was his name') drives it hardest, near +10. Quiz-style 'what is the name of' questions and requests for…","st":"tested","i":9,"c":8,"m":[["name of place",4.8],["word",3.2],["title",3.2],["category: city",1.6],["name",1.5],["price (control)",1.0],["category: cuisine",1.0],["address",0.7],["date (control)",0.4],["weather (control)",-0.2]]},{"id":"L27_E23_N290","k":"m","L":27,"E":23,"N":290,"t":"A neuron for the closing backtick of a newly named code identifier","l":"This neuron marks the moment a quoted code name closes in a 'the ___' slot, the backtick in 'the `sha256sum` function', where the model expects a word like function or option next. It fires hardest when the sentence explains code shown just above. Quoted interface labels fire…","st":"tested","i":5,"c":8,"m":[["backticked name",6.2],["UI label",4.4],["quoted command",2.9],["snake_case name",2.5],["math variable",1.4],["quoted value",0.7],["parenthesized",0.2],["unquoted name",0.0],["plain noun",-0.0],["concept",-0.5]]},{"id":"L27_E37_N501","k":"m","L":27,"E":37,"N":501,"t":"A neuron for 'now do the same to the rest'","l":"This neuron tracks a common move in instructions and explanations: an action has just been done to one member of a set, and the text is about to do the same to the next one or the rest. It peaks on the slot before 'the remaining' or 'another', as in 'Repeat with the remaining…","st":"tested","i":7,"c":8,"m":[["then same",1.4],["do the same",1.2],["repeat",1.1],["continue",1.0],["one-off",0.7],["decorate (control)",0.4],["serve (control)",-0.0],["in common",-0.0],["pairs (control)",-0.1]]},{"id":"L27_E40_N114","k":"m","L":27,"E":40,"N":114,"t":"A negative neuron for whether colors, clothes and food go together","l":"This neuron goes strongly negative when a text asks or states whether two things go together, especially colors, clothes, furnishings or food. The dip lands on the pairing verb ('Does gray match', 'Which colors go') and is deepest in dictionary definitions of 'go with = look…","st":"tested","i":7,"c":8,"m":[["apps (non-aesthetic)",-0.0],["buses (control)",-0.0],["trains (control)",0.0],["ties",-3.5],["shoes",-4.0],["shirts",-4.5],["side dishes",-4.9],["wines",-6.8],["colors",-8.8]]},{"id":"L27_E54_N62","k":"m","L":27,"E":54,"N":62,"t":"A neuron for 'where appropriate' in a text's notes on itself","l":"This neuron fires negatively when a text says that something it does applies only in some cases, as in 'Where appropriate, we have listed…' or 'When applicable, I…'. The deepest firing lands on the qualifier and on the subject right after it, reaching about −7. Positive firings…","st":"tested","i":5,"c":6,"m":[["if necessary",-0.0],["always (control)",-0.0],["yesterday (control)",-0.0],["below (control)",-0.1],["often",-2.5],["wherever possible",-2.9],["where relevant",-3.6],["when applicable",-3.7],["some cases",-3.8],["where appropriate",-4.1]]},{"id":"L27_E55_N184","k":"m","L":27,"E":55,"N":184,"t":"A neuron that recognizes the PyTorch char-RNN tutorial's training loop","l":"Inside an expert that takes in machine-learning code and docs, this neuron fires on passages that look like specific well-known tutorials. Its clearest trigger is the loop opener of the PyTorch char-RNN training function (about +2.7). It also fires on the count noun in…","st":"tested","i":5,"c":5,"m":[["augment args",3.0],["model+opt only",2.7],["char-RNN full",2.6],["short generic",2.6],["quicksort control",2.5],["tensors only",2.4],["fruit control",2.2],["seq2seq variant",1.7],["generic ML",1.6]]},{"id":"L27_E81_N264","k":"m","L":27,"E":81,"N":264,"t":"The period after 'this is what every body normally does'","l":"This neuron marks the end of a sentence that states the ordinary baseline: what every body, organism or machine routinely has or does, such as 'Everyone has bacteria living on their skin.' It fires on the closing period or comma, mostly in health prose but not limited to it.…","st":"tested","i":5,"c":6,"m":[["machine (halo)",3.2],["universal baseline",2.8],["body response",1.6],["everyday universal",1.6],["rare case",0.6],["natural process",0.5],["dated event",-0.0],["instruction",0.0],["should-claim",-0.6],["usual duration",-1.2]]},{"id":"L27_E90_N397","k":"m","L":27,"E":90,"N":397,"t":"A signed dial for the fixed later versus earlier neighbor","l":"Inside step-by-step reasoning about ordered things like intervals, arrays, search-tree keys and adjacent words, this neuron reads which side an already-fixed reference sits on. It turns positive when the current item is checked against a settled later neighbor, such as the next…","st":"tested","i":7,"c":8,"m":[["next existing",0.8],["successor",0.7],["right neighbor",0.6],["control",0.6],["control",-0.0],["subsequent (halo)",-0.1],["following",-0.3],["previous",-1.9],["left half",-2.0],["preceding",-3.0]]},{"id":"L28_E120_N233","k":"m","L":28,"E":120,"N":233,"t":"A signed marker for the second half of a pair","l":"This neuron tracks the second half of a two-part parallel, with a sign. Inside a second copy of a phrase, it fires on the word that swaps in the counterpart of the first copy's word, such as 'small' after 'large'. It goes strongly negative on the label or connective that opens…","st":"tested","i":6,"c":8,"m":[["small (antonym)",1.0],["tiny",1.0],["medium",1.0],["'a' (lead-in)",1.0],["purple (no contrast)",1.0],["large (repeat)",1.0],["unrelated",0.0],["'compared'",0.0],["Option B",-0.3]]},{"id":"L28_E120_N51","k":"m","L":28,"E":120,"N":51,"t":"A neuron that flags the next entry after square-root side lengths","l":"This neuron tracks lists of computed lengths in worked math. Along a list of side lengths or distances whose earlier entries came out as square roots, it fires on the colon or equals sign that opens the next entry, most strongly when the coming value is a plain number, and it…","st":"tested","i":5,"c":6,"m":[["plain distance",2.4],["side label",2.4],["norm form",1.7],["equals form",1.3],["range 'to'",0.3],["rule comma",-0.0],["prose control",-0.0],["non-math label",-0.0],["term list",-0.1],["point placement",-0.8]]},{"id":"L28_E13_N204","k":"m","L":28,"E":13,"N":204,"t":"A signed name-spelling lookup: Men|k up, K|ask down","l":"Inside the second piece of rare, split-up names, this neuron works as a signed spelling lookup. The 'k' in Men|k (as in Menkara) pushes it strongly positive. The 'ask' after K (as in Kaskautas) pushes it strongly negative. Most other name pieces and ordinary words leave it near…","st":"tested","i":3,"c":8,"m":[["Men|k (core)",6.7],["Y|ng",1.4],["Ken|k",0.8],["Hen|k",0.2],["control",-0.0],["control",-0.0],["N|ask",-1.5],["K|usk",-2.3],["K|ask (pole)",-11.6]]},{"id":"L28_E140_N67","k":"m","L":28,"E":140,"N":67,"t":"Predicting the next gentle example: walking, folding laundry, a whisper","l":"This neuron fires at the moment a text is about to give another example of something gentle and everyday: a light exercise, an easy chore, or a quiet sound. Two cues add up: a word naming the low level ('light', 'moderate', a 50-decibel reading) and gentle items already in the…","st":"tested","i":6,"c":7,"m":[["easy chore",5.1],["gentle activity",4.9],["light cardio",4.0],["calm pastime",3.5],["off-topic",3.5],["period chore",2.6],["vigorous",2.1],["vigorous",1.8],["gentle stretch",1.7]]},{"id":"L28_E143_N432","k":"m","L":28,"E":143,"N":432,"t":"A neuron that signs products: hard-use gear positive, airbeds negative","l":"Inside product copy, this neuron fires at the spot where the next spec value or material is about to be named, and its sign depends on what the product is. Hoses, dive masks, tires and protective gear push it up. Airbeds, pool floats, toys and souvenir or made-in-America gift…","st":"tested","i":5,"c":6,"m":[["air hose",5.4],["dive mask",2.2],["garden hose",2.2],["truck tire",1.9],["phone (control)",1.3],["brake pads",1.0],["novel (control)",0.9],["tent",-1.0],["pool float",-1.4],["airbed",-2.7]]},{"id":"L28_E2_N287","k":"m","L":28,"E":2,"N":287,"t":"A neuron that fires when a lesson about folly is coming","l":"This neuron waits for a lesson. When earlier text frames a bad outcome as self-inflicted ('that's what I get', 'a cautionary tale of what happens', 'the dangers of'), it fires on the word that opens the slot for the folly, before the folly is named and whatever it turns out to…","st":"tested","i":7,"c":8,"m":[["neutral example",3.8],["sad example",3.3],["cautionary tale",3.1],["harsh lesson",3.1],["painful reminder",3.0],["statistics",3.0],["delightful (control)",2.7],["technical analysis",1.5],["happy (control)",0.0]]},{"id":"L28_E30_N89","k":"m","L":28,"E":30,"N":89,"t":"A neuron that writes 'location' and reacts weakly to where-phrases","l":"This neuron sits in an expert that is only chosen for place text, and it reads one signed axis about locating things. Its strong negative pole fires where location itself is named, as in 'the island's strategic location' or 'Situated on the Ring of Fire'. Pushing it negative…","st":"tested","i":5,"c":6,"m":[["control: tourism",0.2],["control: economy",0.1],["valley (where)",-1.0],["situated",-1.5],["offset",-2.5],["position",-6.1],["strategic location",-7.5],["remote location",-8.7],["location",-8.9]]},{"id":"L28_E40_N257","k":"m","L":28,"E":40,"N":257,"t":"Name fragments read as known entities or as ordinary words","l":"This neuron reads the odd, non-word fragments inside split-up names and codes, such as the hn in ca.uhn or the id in Ghidra, and asks whether the context makes that fragment part of a specific known entity. A disassembly sentence pushes Ghidra's piece strongly positive. A domain…","st":"tested","i":4,"c":6,"m":[["known tool",0.0],["known tool",0.0],["related tool",0.0],["wrong entity",-0.0],["word-like name",-0.0],["control",0.0],["control",0.0],["misspelled",-0.1],["word compound",-0.1]]},{"id":"L28_E70_N482","k":"m","L":28,"E":70,"N":482,"t":"A music neuron that likes definitions and dislikes lists","l":"Inside music text this neuron reads the sentence slot rather than the topic. It goes positive at the connector that ends a definition of a musical concept, peaking at 'syncopation is a disruption of', at about +2.3. It goes negative at the commas and 'and' tokens of lists of…","st":"tested","i":4,"c":6,"m":[["syncopation",2.3],["hemiola",1.4],["dissonance",0.6],["cadence",0.2],["techno",0.2],["meter",-0.0],["control: heart",0.0],["control: traffic",0.0],["dubstep",-0.3],["polyrhythm",-0.9]]},{"id":"L29_E117_N195","k":"m","L":29,"E":117,"N":195,"t":"Citing the rule versus naming the hazard in standards text","l":"Inside standards and regulatory text, this neuron separates how a rule is identified from what it regulates. It fires positive on a cited document's designation, the slot just after an issuing body's acronym, as in 'API RP 2SK'. It fires strongly negative on the regulated…","st":"tested","i":4,"c":6,"m":[["API bulletin",3.4],["API RP code",3.0],["NORSOK code",1.4],["ISO code",1.3],["fake body",1.0],["model code",0.9],["control: everyday",-0.2],["control: person",-1.0],["class tier",-2.0],["hazard class",-2.3]]},{"id":"L29_E140_N249","k":"m","L":29,"E":140,"N":249,"t":"A neuron that knows when object files are about to be linked","l":"This neuron tracks native build state: whether text is in the middle of compiling source to object files and linking them. It fires hardest just before an object file or a link target is named, such as a Makefile link-rule colon or the first .o in a link command, and it grows…","st":"tested","i":7,"c":4,"m":[["g++ -c",8.9],["gcc -c",8.2],["archive",8.0],["clang -c",7.3],["pandoc",5.4],["-S assembly",5.2],["nasm object",4.9],["recipe",4.9],["edit source",3.1]]},{"id":"L29_E196_N34","k":"m","L":29,"E":196,"N":34,"t":"A neuron for unpacking several values, silenced by textbook helpers","l":"In numeric code, this neuron tracks a moment of binding several values at once. It fires on the comma after the first name in an unpacking like next_state, reward, done = env.step(...). It fires more weakly in loop-variable tuples and in brute-force loops just after a 'what's…","st":"tested","i":4,"c":6,"m":[["unpack comma",3.1],["second comma",3.1],["loop target",3.0],["short target",2.7],["argument comma",0.1],["string build",0.1],["comment",-0.0],["remainder indent",-0.3],["lerp pair",-0.5]]},{"id":"L29_E1_N300","k":"m","L":29,"E":1,"N":300,"t":"A code-prose dial: container operations up, construct praise down","l":"Inside prose that explains code, this neuron works as a signed dial. It turns positive at the connective that follows a completed description of what a method does to a collection, as in 'returns the smallest element in the stack[ in] O(1)'. It turns more strongly negative when…","st":"tested","i":5,"c":6,"m":[["max query",3.8],["extremum query",3.0],["remove op",1.3],["append op",1.0],["control: non-code",0.6],["control: authorship",-0.3],["write (halo)",-0.8],["abstraction",-2.2],["praise",-3.8]]},{"id":"L29_E200_N443","k":"m","L":29,"E":200,"N":443,"t":"A neuron for the teacher who says 'I'm learning too'","l":"This neuron tracks who is doing the learning. It fires when a teacher, parent or other capable adult steps in as a fellow beginner, as in 'I'm learning this too', reaching about +10 on 'learning'. It stays quiet when that adult speaks as an expert. It turns negative in…","st":"tested","i":8,"c":6,"m":[["no 'learn'",9.6],["co-learner",9.4],["together",8.4],["solo learning",8.0],["find out",3.2],["authority",2.7],["control",2.7],["joins in",1.4],["control",0.2],["roadmap",-2.6]]},{"id":"L29_E215_N305","k":"m","L":29,"E":215,"N":305,"t":"Raw base64 blobs push it up, hex and digit IDs down","l":"This neuron is a payload detector for raw base64. It fires deep inside free-standing encoded blobs such as image data URIs, certificates and compressed data, and the evidence (length, mixed case, '+', known headers) builds up over the run. It turns negative inside hexadecimal…","st":"tested","i":4,"c":8,"m":[["JPEG base64",1.3],["PNG base64",0.9],["certificate",0.6],["prose",0.1],["encoded English",-0.2],["number array",-0.3],["hex hash",-0.3],["digit ID",-0.3],["no-plus base64",-0.4],["JWT",-0.4]]},{"id":"L29_E218_N307","k":"m","L":29,"E":218,"N":307,"t":"Derived 'would' versus guessed 'must': a dial for claim stance","l":"This neuron reads how a writer stands behind a claim, at the first verb or modal of that claim. It turns positive when the claim is worked out from stated premises about a constrained object, as in 'So a/b would not be in lowest terms.' It turns about three times more strongly…","st":"tested","i":7,"c":6,"m":[["would (core)",2.3],["should",2.1],["plain is",1.4],["likely",1.4],["probably",1.4],["might",1.3],["can",0.9],["control",0.8],["derivation must",0.5],["must surely",0.5]]},{"id":"L29_E227_N130","k":"m","L":29,"E":227,"N":130,"t":"Liability triggers up, sovereign 'may by law' powers down","l":"In legal text, this neuron's sign tracks what kind of clause is being written. It turns positive on the word right after a fact that triggers a rule, such as 'liable where a third party uses a sign' or 'in respect of which a declaration'. It turns negative when the clause says…","st":"tested","i":5,"c":5,"m":[["third-party trigger",0.9],["liability trigger",0.8],["applicability",0.8],["generic legal",0.6],["exception",0.4],["cooking control",0.2],["legislature power",0.1],["State permission",-0.0],["sport control",0.0],["laws providing for",-0.6]]},{"id":"L29_E240_N493","k":"m","L":29,"E":240,"N":493,"t":"Effect versus mechanism: a signed neuron for ice, water and fire","l":"Inside explanations of what ice, water, fire and heat do, this neuron changes sign with the part of the explanation. It is positive on the word stating a material's effect ('can help', 'serve to', 'source of'). It is strongly negative just before the text explains the mechanism…","st":"tested","i":6,"c":6,"m":[["social ctrl",0.0],["decay",-0.2],["towel smells",-0.5],["mechanical ctrl",-0.7],["compost heat",-1.8],["fire warms",-2.0],["ice cools",-4.2],["towel cools",-4.2],["sweat cools",-4.9]]},{"id":"L29_E249_N301","k":"m","L":29,"E":249,"N":301,"t":"A hinge-before-the-slot detector for a few familiar problem families","l":"This neuron marks the word just before an answer slot, such as the '=' in 'evaluate this polynomial at x =' or the 'by' in 'divisible by'. It does this only inside a few familiar kinds of text: polynomial-evaluation and divisibility problems, geometry incidence problems, and…","st":"tested","i":3,"c":4,"m":[["long poly eval",3.4],["value of poly",3.0],["cubic eval",2.6],["short poly",2.3],["continuity",1.0],["derivative",0.0],["control: weather",0.0],["control: novel",0.0],["design pattern",-0.8]]},{"id":"L29_E24_N400","k":"m","L":29,"E":24,"N":400,"t":"A Chinese-answer junction neuron: programming openers up, formula steps down","l":"This neuron reads junction words in Chinese assistant answers, and it is the only kind of text its expert receives. It turns positive at the close of an answer's opening frame, such as 在Python中 after a Python how-to question. It turns sharply negative, to about −5, on the…","st":"tested","i":4,"c":5,"m":[["在Python中",1.4],["opener comma",1.4],["使用…时",0.6],["function intro",0.1],["wrong language",-0.0],["control",0.0],["control noun",0.0],["definition",-0.1],["formula 再",-4.9],["formula 然后",-5.7]]},{"id":"L29_E98_N24","k":"m","L":29,"E":98,"N":24,"t":"A code neuron that anticipates shuffling a freshly built list","l":"This neuron tracks random operations on collections in code. It turns mildly positive while a random draw is being set up, such as `import random` or the `len` bound of `randint`. It goes far more negative when a list has just been built and the model expects it to be shuffled…","st":"tested","i":9,"c":6,"m":[["prices",0.3],["user names",-0.1],["scrambled literal",-2.0],["split words",-2.0],["ordering comment",-6.6],["ticket range",-10.9],["shuffle comment",-11.0],["range 1..40",-12.7],["Array.from range",-12.8]]},{"id":"L2_E169_N134","k":"m","L":2,"E":169,"N":134,"t":"A detector-noun reader that turns negative on brand tails","l":"This early-layer neuron reads the second word of short technical compounds. When 'detector' or 'detection' follows a modifier like gas, data or object, it fires strongly. When the same slot holds a brand or name tail, as in Gas|quet or Data|Camp, it turns negative. The router…","st":"tested","i":3,"c":7,"m":[["detector",2.4],["plural",2.4],["detection",2.1],["analyzer",0.9],["sensor",0.6],["control",0.2],["meter",0.1],["Gas|oline",-0.0],["control",-0.2],["Gas|quet name",-0.7]]},{"id":"L2_E213_N184","k":"m","L":2,"E":213,"N":184,"t":"A neuron that memorizes two short lists of rare words","l":"This neuron watches for a short memorized list of rare words that the tokenizer splits into pieces. It responds on the piece that completes the word. One list, including Turkish kurtar-, Latin Deus and 'Crazy Blues', drives it positive. A second list of rare names, including…","st":"tested","i":3,"c":2,"m":[["Deus (core)",2.1],["-aton (core)",1.9],["Blues (core)",1.3],["term (halo)",0.5],["Dios (variant)",0.4],["plain name",-0.0],["plain words",-0.0],["-pet (neg)",-0.8],["Tetra Tech (neg)",-1.4],["Bechterew (neg)",-1.7]]},{"id":"L2_E221_N59","k":"m","L":2,"E":221,"N":59,"t":"A neuron for the UN as an institution: 'United Nations System'","l":"This early-layer neuron marks the moment a word names an organization as a whole or one of its parts, right after the organization's name: most of all ' System' in 'United Nations System', and more weakly UNCG, UNESCO Collection or 奉军. The neuron itself responds to almost any…","st":"tested","i":4,"c":6,"m":[["UN System",2.1],["acronym core",2.1],["UN body",1.4],["variant name",1.0],["unit noun",0.8],["weak head",0.3],["not routed",-0.0],["generic control",-0.0],["glued negative",-1.4]]},{"id":"L2_E237_N274","k":"m","L":2,"E":237,"N":274,"t":"A lookup neuron for 'drive', delivered after 'at the'","l":"This neuron mostly looks up one word: the noun 'drive'. The router hands it the word mainly right after 'at the' or 'to the', or inside road names, so its strongest hits are 'at the drive-in' and 'Atlantic Drive' (around +2 to +4). A few other words in the same frame, such as…","st":"tested","i":3,"c":8,"m":[["Drive (caps)",4.2],["drive-in",2.9],["drive-thru",2.9],["register",0.6],["Halfway",0.4],["diner (ctrl)",0.1],["garage (ctrl)",0.1],["hotel (ctrl)",-0.0],["Fraser",-0.0],["Air Base",-0.3]]},{"id":"L2_E246_N331","k":"m","L":2,"E":246,"N":331,"t":"A neuron that recognizes the spelled name Antioch in any script","l":"This neuron is a spelling detector for one name. It fires on the final piece of Antioch or Antiochus, in Seleucid king lists, Bible dictionaries, modern news, or Chinese, Japanese and Russian transliterations, whenever the An-ti-o-k sound sequence is present. Most of its…","st":"tested","i":5,"c":8,"m":[["Antioch",2.7],["Latin form",2.1],["partial stem",2.1],["sibling city",1.1],["look-alike",0.2],["unrouted sibling",-0.0],["control",0.0],["control",0.0],["Mc surname",0.0]]},{"id":"L2_E253_N507","k":"m","L":2,"E":253,"N":507,"t":"A 'myth' and 'Angels' word detector that dips on grundlegend","l":"This neuron is a word detector, not a topic detector. It fires on the word 'myth', in any spelling or split, and on 'Angels'. It turns negative on the Germanic stem 'leg', and most sharply on the 'leg' in German and Scandinavian 'grundlegend', where the negative firing helps the…","st":"tested","i":3,"c":7,"m":[["angels",1.2],["myth (core)",0.9],["plural myth",0.9],["mythology",0.9],["sacred halo",0.8],["control",0.2],["control",0.1],["control",-0.0],["basic",-0.3],["leg stem",-0.6]]},{"id":"L2_E30_N122","k":"m","L":2,"E":30,"N":122,"t":"A signed pair lookup inside rare names: Nk|umb up, Fl|are down","l":"This layer-2 neuron behaves like a small signed lookup table over pairs of word pieces inside rare names and identifiers. It fires strongly on 'umb' after 'Nk' in Nkumbula, and swings negative on 'are' after 'Fl' in Flare. The same piece can flip sign depending on the piece…","st":"tested","i":3,"c":6,"m":[["Nk|umb core",2.8],["k|umb variant",2.3],["wan|aw",1.9],["surname Mun",1.6],["made-up aw",1.5],["Nd|umb halo",0.9],["Bard negative",-0.0],["control",-0.0],["plain name",-0.0],["Burn negative",-1.0]]},{"id":"L2_E52_N34","k":"m","L":2,"E":52,"N":34,"t":"A neuron that fires on 'first cousin', 'first floor', 'first order'","l":"This layer-2 neuron is essentially a negative detector for a small set of nouns, order/derivative, cousin and floor, when 'first' (or primo, erster, premier) comes just before them. The router agrees: after 'first' its expert ranks at the top, while after 'second' or 'distant'…","st":"tested","i":4,"c":8,"m":[["interrupted",-0.0],["distant",-0.0],["bare noun",-0.0],["control",-0.0],["first + other",-0.1],["second",-1.6],["plural",-1.7],["1st cousin",-2.1],["first cousin",-2.4]]},{"id":"L2_E99_N215","k":"m","L":2,"E":99,"N":215,"t":"A signed word lookup, unlocked by 'Open': Solar up, UNIX down","l":"This neuron works like a small lookup table on the current word. 'Solar' is strongly positive, a few name pieces are mildly positive, and words like UNIX, Heaven, lock and Internet are negative. The router decides which values actually reach the output. It mostly admits the word…","st":"tested","i":3,"c":5,"m":[["Solar (core)",5.2],["Sun",1.1],["College (halo)",0.4],["Door (control)",-0.0],["Source (control)",-0.0],["Stack (control)",0.0],["Comput",-1.3],["Heaven",-1.3],["Internet",-2.0],["UNIX",-3.2]]},{"id":"L30_E104_N270","k":"m","L":30,"E":104,"N":270,"t":"Name tokens signed by field: wildlife positive, crop farming negative","l":"This neuron responds to pieces of people's names, such as initials and surname onsets, and its sign depends on the field of the surrounding text. Names in writing about wild animals and their habitats push it positive. The same kinds of names in crop and farm-economy writing…","st":"tested","i":6,"c":8,"m":[["timber/wildlife",3.3],["plover nests",3.2],["trout stream",2.0],["mussel survey",1.8],["physics",-0.1],["database",-0.2],["wetland/farm",-0.6],["cattle pasture",-0.9],["corn yield",-2.5],["weed science",-3.7]]},{"id":"L30_E111_N452","k":"m","L":30,"E":111,"N":452,"t":"An Italian-expert neuron that prefers Machiavelli to Newton's laws","l":"Inside an expert the router reserves for Italian, this neuron tracks register more than language. It lights up on the little words and word pieces of literary or archaic Italian, such as Machiavelli or opera libretti, and stays quiet on news and chat Italian. It turns negative…","st":"tested","i":5,"c":5,"m":[["Machiavelli",2.0],["imprints",1.8],["Guicciardini",0.9],["modern paraphrase",0.2],["recipe control",0.1],["parliament control",0.1],["Kepler",-0.5],["Newton",-0.5],["sacraments",-1.0]]},{"id":"L30_E126_N471","k":"m","L":30,"E":126,"N":471,"t":"A neuron for 'now, what could be the question?'","l":"This neuron responds to a specific instruction-dataset move: after material has already been shown, such as an answer key, a claim or a rationale, the prompt asks for the question or verdict behind it ('Now, what could be the question?'). Verbatim templates reach about +5, while…","st":"tested","i":4,"c":6,"m":[["reverse question",5.4],["short reverse",2.5],["text-to-response",1.7],["claim verdict",1.5],["rationale request",1.3],["translate",-0.0],["plain prose",0.0],["off-topic prose",0.0],["format boilerplate",-0.2],["review byline",-1.8]]},{"id":"L30_E143_N199","k":"m","L":30,"E":143,"N":199,"t":"Surname terms push it up, common-word rule names push it down","l":"This neuron sorts the modifier word of a fixed technical name by what kind of word it is. When a person's surname serves as the term (Raynaud's phenomenon, Taylor series, Krebs cycle), the neuron fires positive, peaking near +9 on Raynaud in clinical notes. When an ordinary word…","st":"tested","i":7,"c":8,"m":[["Bessel (core)",3.6],["Taylor (core)",3.4],["Fourier (halo)",2.7],["Cauchy (core)",1.4],["invented name",0.3],["control",-0.0],["control",0.0],["shoelace (rule)",-0.9],["Law (of Sines)",-0.9],["distance (rule)",-1.7]]},{"id":"L30_E180_N395","k":"m","L":30,"E":180,"N":395,"t":"A neuron for bracket fragments quoted as data, not code","l":"This neuron responds to short symbol strings that are mentioned rather than used. It fires most on bracket fragments like (() quoted as test cases, and less on drill operators, cube moves, newly introduced math variables and the subjects of claims to be checked. Brackets used…","st":"tested","i":5,"c":6,"m":[["(() core",4.4],["mixed ({",3.8],["()()",3.1],["((",2.4],[")))",1.9],["letters xq",0.9],["word banana",0.3],["greeting (control)",0.1],["phrase (control)",-0.5]]},{"id":"L30_E218_N268","k":"m","L":30,"E":218,"N":268,"t":"A neuron that anticipates the next winding step in coil devices","l":"This neuron tracks one moment in technical writing: a wire-on-core device such as a transformer, inductor or shake flashlight is being built or specified, and the next winding detail is about to be stated. It fires on the 'and', 'through', 'number of' or blank before a turn…","st":"tested","i":5,"c":7,"m":[["transformer",14.1],["step-up",11.7],["motor",7.0],["solenoid",5.4],["inductor",5.3],["loudspeaker",3.4],["bicycle",-0.0],["resistor",-0.1],["lemon battery",-0.4]]},{"id":"L30_E228_N395","k":"m","L":30,"E":228,"N":395,"t":"A neuron for why lenders must be paid a premium","l":"This neuron fires strongly negative at the moment a text explains why a lender, bondholder or investor must be paid extra: inflation erodes repayment, a borrower might default, a bond runs longer. Costs folded into a shop's price fire it more weakly. Plain price changes and…","st":"tested","i":5,"c":7,"m":[["neutral control",-2.0],["plain price rise",-2.5],["insurance risk",-2.6],["cost pass-through",-4.4],["risk, no money",-5.1],["bond duration",-6.6],["volatility only",-7.0],["inflation erosion",-9.5],["inflation + default",-10.2]]},{"id":"L30_E24_N363","k":"m","L":30,"E":24,"N":363,"t":"A Spanish-expert neuron that recalls BMI, proverbs, and María Zambrano","l":"This neuron belongs to an expert the model routes Spanish text to, and on most tokens it does almost nothing. It speaks up on a handful of passages that recur in Spanish instruction data. It goes positive at a specific word inside the BMI formula 'IMC = peso / altura²', the…","st":"tested","i":4,"c":6,"m":[["BMI formula",1.3],["requirement list",1.3],["proverb",1.0],["area formula",0.3],["English BMI",0.3],["velocity (control)",0.1],["other phrase",-0.0],["Greene dates",-0.0],["definition",-0.2],["Zambrano",-2.3]]},{"id":"L30_E7_N365","k":"m","L":30,"E":7,"N":365,"t":"Lime holds the pH, polymers hold the charge: a signed dial","l":"Inside a chemistry-focused expert, this neuron works as a signed dial for acid-base talk. It turns positive when an agent, usually an alkali such as lime or ammonia, is being applied to set or hold acidity in a recipe or process. It turns negative when charge or acidity is a…","st":"tested","i":5,"c":6,"m":[["lime",2.1],["ammonia",1.6],["caustic soda",1.5],["acid",1.2],["vinegar",1.1],["stirring (control)",0.8],["gravel (control)",0.2],["salt (control)",-0.5],["charged resin",-1.6]]},{"id":"L30_E82_N102","k":"m","L":30,"E":82,"N":102,"t":"A neuron that tracks how far along the alphabet a sorted list is","l":"This neuron tracks position in alphabetically sorted lists. When items share a long repeated lead-in, such as 'Fellows of the' or a dotted Java package path, it fires negative at the first token where the new item diverges. The value grows with how late in the alphabet that item…","st":"tested","i":9,"c":6,"m":[["Ze",-0.0],["Wo",-0.0],["Tu",-0.0],["Se",-0.0],["Mo (mid)",-0.0],["Ge (early)",-0.0],["Ac (out of order)",-0.0],["repeat (control)",-0.1],["Zo (core)",-5.5]]},{"id":"L30_E82_N164","k":"m","L":30,"E":82,"N":164,"t":"A neuron that counts repeated heads in index lists","l":"This neuron tracks index-style lists. When a new entry opens with the same word as the entries above it, and those entries branch differently after it, it fires on that shared head word. The signal climbs with each repeat, from zero to about +3 over a dozen lines. Its opposite…","st":"tested","i":4,"c":6,"m":[["shared head",3.2],["head + 2nd word",3.2],["new branch",3.2],["head variant",0.8],["function word",0.1],["2nd word only",-0.0],["new head",-0.1],["measurement",-0.1],["table cells",-0.3]]},{"id":"L30_E90_N450","k":"m","L":30,"E":90,"N":450,"t":"A neuron for 'Unlike X, however': limiting what was just claimed","l":"This neuron marks the point where a writer turns back to limit something they just said. It fires on a fronted phrase such as 'Unlike a ribbon' or 'Even this project' when that phrase revisits the subject of a claim made moments earlier, and it pushes the model toward 'however'.…","st":"tested","i":6,"c":8,"m":[["Unlike + comparand",12.8],["Even + comparand",9.9],["variant article",6.4],["shifted referent",2.4],["similarity only",1.0],["control",0.1],["unrelated referent",0.0],["new alternative",-0.0],["control",-0.1]]},{"id":"L30_E92_N352","k":"m","L":30,"E":92,"N":352,"t":"A code neuron that separates value names from label names","l":"Inside code, this neuron reads what a name is about as the name is being spelled out. It turns on for names that hold a quantity or a literal value: distances, sizes, colours, regex patterns, and especially members of a palette or constants block. It turns negative for names…","st":"tested","i":6,"c":8,"m":[["physical constant",3.6],["size",2.6],["colour",2.5],["style",2.2],["generic (halo)",1.5],["pattern",1.4],["control: function",0.1],["control: import",0.0],["status",-0.4],["opaque key",-2.5]]},{"id":"L31_E104_N244","k":"m","L":31,"E":104,"N":244,"t":"A neuron for 'this problem trains the students' skills'","l":"This neuron tracks talk about exercising someone's abilities, especially the stock line in Chinese-style math solutions: 'This problem trains/tests the students' skills.' It fires on the token that opens the skill object, either the verb ('trains') or the learner ('students'…","st":"tested","i":4,"c":4,"m":[["student",5.8],["students",5.6],["candidates",4.8],["your",4.6],["players",4.1],["ability (no person)",3.0],["computer",2.7],["function",2.7],["weather (control)",2.7],["bridge (control)",2.7]]},{"id":"L31_E105_N72","k":"m","L":31,"E":105,"N":72,"t":"A PCIe-flag neuron that nudges lspci dumps toward '-'","l":"This neuron is a negative-valued detector for PCIe register dumps. Inside lspci -vv capability and status lines it grows steadily negative and is strongest at the flag slot, such as ' Enable' in 'MSI-X: Enable+'. There it tilts the next token toward '-' rather than '+'. Zeroing…","st":"tested","i":6,"c":6,"m":[["version string",3.0],["non-flag cap",2.3],["GPU spec",-1.7],["ethtool flag",-2.8],["garden mimic",-2.8],["cpuinfo",-3.7],["Status flags",-5.6],["LnkCtl sign",-6.0],["MSI-X Enable",-9.2],["MSI Enable",-9.6]]},{"id":"L31_E161_N369","k":"m","L":31,"E":161,"N":369,"t":"A neuron that asks whether free riders can be charged","l":"This neuron asks one question about a shared good: can the people who benefit be charged or kept out? It swings strongly negative, down to about -18, at the moment a text poses that question, as with the lighthouse owner who 'cannot...'. It does so whether the answer is yes or…","st":"tested","i":7,"c":6,"m":[["toothbrush",-12.4],["sandwich",-13.6],["locked library",-13.8],["bakery",-14.7],["fenced garden",-14.8],["lighthouse",-14.9],["fireworks",-15.3],["streetlight",-15.4],["shoveled sidewalk",-17.3]]},{"id":"L31_E167_N396","k":"m","L":31,"E":167,"N":396,"t":"Opening the next worded label in a code or figure series","l":"This neuron fires on the token that opens a short worded label in technical text, such as the '#' before the next section comment in code or the number in the next figure caption. It fires strongly only after earlier items have set up an item-then-label rhythm, peaking around…","st":"tested","i":4,"c":6,"m":[["Spanish comment",4.7],["Portuguese comment",4.2],["English comment",3.1],["one-word label",3.1],["commented code",3.1],["TODO marker",3.1],["FLAN boilerplate",3.1],["no label",2.7],["docstring",2.4],["docblock",1.8]]},{"id":"L31_E170_N0","k":"m","L":31,"E":170,"N":0,"t":"A neuron that sees an infinite regress coming","l":"This neuron anticipates an infinite regress. It fires at the comma or 'and so' right after an argument shows a link that repeats itself, such as a guard who needs a guard or a sign that needs another sign, before 'ad infinitum' appears. The pattern holds across English, Spanish…","st":"tested","i":9,"c":7,"m":[["homunculus",22.9],["guard regress",20.8],["Peirce semiosis",18.0],["Zeno halving",13.7],["mirrors (halo)",12.7],["recurrence",1.8],["neutral",0.1],["finite list",-0.2],["bounded road",-0.4],["infinity defined",-2.4]]},{"id":"L31_E182_N356","k":"m","L":31,"E":182,"N":356,"t":"Naming what matter is made of, not what's inside atoms","l":"This neuron marks the moment in science writing just before the text names the particles a substance, solution or beam is made of. Examples are 'the mass of one mole of …' and 'a solution which contains …'. Talk about the inside of a single particle, such as energy levels, spin…","st":"tested","i":6,"c":8,"m":[["gas swarm",3.1],["crystal array",2.5],["one mole of",1.7],["dozen (count)",1.3],["control: grading",0.0],["control: history",-0.1],["bond angle",-0.8],["hyperfine states",-1.1],["energy levels",-1.2]]},{"id":"L31_E213_N455","k":"m","L":31,"E":213,"N":455,"t":"Everyday foods and habits as health agents versus clinical drug prose","l":"This neuron tracks what is acting on a body. It fires when lay health, garden or pet advice names an everyday food, plant or habit as the agent, such as grapefruit with statins, garlic against aphids, or 'stop smoking'. It peaks on the agent or just before its effect is…","st":"tested","i":5,"c":8,"m":[["health food",2.9],["habit",1.8],["herb",1.6],["less typical food",1.6],["drink",1.4],["activity control",1.0],["lay drug",0.9],["clinical drug",0.7],["non-body control",0.5]]},{"id":"L31_E241_N27","k":"m","L":31,"E":241,"N":27,"t":"A neuron for where-to-look citation numbers, against reference-list dashes","l":"This neuron marks the head locator of a humanities or scripture citation, the number that says where in a source to look. It fires on a Bible chapter after the book name, on a page number after p. in endnotes, and on the separator after a church or legal document title. It goes…","st":"tested","i":4,"c":6,"m":[["Bible chapter",3.6],["Psalm chapter",1.5],["halo: chapters",0.5],["unexpected chapter",0.4],["title separator",0.2],["control: prose",0.0],["issue date",-0.1],["control: prose year",-0.2],["endnote page",-0.3],["reference dash",-1.5]]},{"id":"L31_E2_N205","k":"m","L":31,"E":2,"N":205,"t":"Measured amounts versus trig angles in coordinate math","l":"This neuron tracks what a number is doing. It fires just before a measured, real-world amount in prose, listings or spec sheets, the kind of number that will take a unit, percent or price. It goes strongly negative on digits in coordinate math. The deepest case is the angle…","st":"tested","i":5,"c":6,"m":[["spec volume",2.4],["decimal value",2.0],["survey percent",1.8],["unitless count",0.8],["config setting",0.6],["year (label)",0.0],["coordinate length",-0.6],["exact algebra",-0.9],["trig angle",-3.6]]},{"id":"L31_E70_N33","k":"m","L":31,"E":70,"N":33,"t":"Counted sets versus 'companies like X': a register dial","l":"This neuron reads the register in which a plural category noun appears. It turns positive when a vague quantifier such as 'several' or 'a number of' marks a set of real members in reference or how-to prose ('a number of rivers'). It turns negative when the noun names a kind to…","st":"tested","i":7,"c":8,"m":[["several + business",1.8],["several (core)",1.6],["a number of",1.6],["many good",1.5],["kind + like",0.2],["exact numeral",0.0],["control: person",0.0],["control: singular",0.0],["bare plural",-0.5],["AI kind",-1.8]]},{"id":"L31_E99_N220","k":"m","L":31,"E":99,"N":220,"t":"A neuron that fires on the spelling 'mono', not its meaning","l":"This neuron is a spelling detector for one token: 'mono'. Whatever the word means (one god, one audio channel, a lake in California, a monkey in Spanish), that token pushes it about twenty times further than anything else. Words that mean 'one' but are spelled differently barely…","st":"tested","i":3,"c":9,"m":[["di- (halo)",0.5],["tri- (halo)",0.3],["monk (lookalike)",0.0],["poly-",-0.0],["control",-0.0],["control",0.0],["uni- ('one')",-1.9],["Mono (caps)",-9.8],["mono",-10.1],["mono- compound",-10.1]]},{"id":"L32_E111_N347","k":"m","L":32,"E":111,"N":347,"t":"Copying from context versus recalling obscure names from memory","l":"This neuron tracks whether the model is copying from its own context or pulling a name from memory. Deep inside text that repeats an earlier stretch word for word, especially long boilerplate repeated many times, it fires strongly positive, up to about +10. On real but obscure…","st":"tested","i":9,"c":6,"m":[["name primed",-1.8],["shuffled words",-1.8],["quote setup",-2.3],["paraphrase",-3.5],["3 prior copies",-3.9],["2 prior copies",-3.9],["1 prior copy",-3.9],["no prefix",-3.9],["unrelated",-4.7]]},{"id":"L32_E134_N510","k":"m","L":32,"E":134,"N":510,"t":"A neuron for conceding a caught mistake, not pre-empting one","l":"This neuron tracks how a speaker responds to an objection. It fires strongly on the first comma or period of an assistant's concession right after a user catches the assistant's own mistake (\"You're right[,]\"), and it does so in many languages. The effect weakens when the error…","st":"tested","i":9,"c":6,"m":[["own bug",4.7],["user's bug",4.6],["own fact error",3.9],["truncated reply",3.2],["coworker's bug",1.9],["no error",1.4],["thanks (control)",0.8],["question (control)",0.5],["refusal pushback",-1.9]]},{"id":"L32_E159_N227","k":"m","L":32,"E":159,"N":227,"t":"Helping someone you could versus being told you should","l":"This neuron sits at the word just before an action is named, such as a modal, 'to', or 'before', and signs it by who the action serves. When the action answers someone's stated need ('I was hoping that you could'), it fires strongly positive, reaching about +4.5. When an…","st":"tested","i":6,"c":7,"m":[["someone could",2.5],["you could",1.9],["you can",1.6],["you would",1.2],["you should",0.3],["you must",0.3],["control: were",0.3],["agency should",0.1],["control: rained",0.0]]},{"id":"L32_E15_N243","k":"m","L":32,"E":15,"N":243,"t":"A bracket neuron that favors definitions and disfavors finishing touches","l":"This neuron fires only on tokens routed to its expert, and the router sends tokens there mostly inside an open bracket. Within the bracket, the neuron is positive when the content describes what the just-named term is: its head noun, an alternative name, or a defining 'that is…","st":"tested","i":4,"c":5,"m":[["defining verb",0.9],["aside onset",0.3],["restatement",0.2],["alt name",0.0],["or-name",0.0],["head NP",0.0],["eventive",0.0],["control: locator",0.0],["year tail",-0.0],["control: ref",-0.2]]},{"id":"L32_E178_N91","k":"m","L":32,"E":178,"N":91,"t":"A lead-in neuron for general principles of care","l":"This neuron marks the moment just before a text lays out general principles of treatment, care or conduct. It fires on the small lead-in word: the comma before 'treatment' in a disease outline, or 'some' in 'once your doctor clears you, here are some guidelines'. What drives it…","st":"tested","i":6,"c":6,"m":[["disclaimer",5.2],["doctor approves",3.3],["surgeon clears",2.8],["plain 'Sure!'",2.0],["PT okay",1.8],["no-vendor refusal",0.7],["off-topic",0.5],["weather control",-0.2],["no-clinic refusal",-0.8]]},{"id":"L32_E194_N341","k":"m","L":32,"E":194,"N":341,"t":"A 'depends on' neuron that separates principles from everyday advice","l":"This neuron fires on the small word linking a property to what it depends on: 'regardless of', 'independent of', 'depends on'. Its sign depends on what is said to depend. Abstract or formal subjects, such as logical validity, sacraments, grammar rules or the speed of light…","st":"tested","i":5,"c":7,"m":[["logic, denied",13.1],["theology, denied",10.5],["logic",4.3],["physics",1.1],["control",-0.0],["control",-0.0],["household",-0.9],["algorithm",-1.1],["pet care",-1.6],["medical",-3.0]]},{"id":"L32_E198_N119","k":"m","L":32,"E":198,"N":119,"t":"A neuron that writes 'will' when a spec prescribes the next step","l":"This neuron reads technical specifications, exercises, contracts and recipes. It fires where the text is about to state the next prescribed step in the future tense: 'The chains will…', 'La función devolverá…', 'Tu rédigeras…'. Pushing it makes the model more likely to write the…","st":"tested","i":7,"c":7,"m":[["will (core)",9.5],["shall",5.4],["must",2.8],["present are",1.9],["may",1.9],["past were",0.6],["conjunction",0.4],["should",0.0],["past verb",-0.0]]},{"id":"L32_E242_N404","k":"m","L":32,"E":242,"N":404,"t":"After inspecting the organ, the neuron anticipates taking tissue","l":"This neuron anticipates a biopsy. Once a doctor has looked inside an organ, usually through a scope, it fires on the words that lead into taking tissue, such as 'may [take]' or 'a number [of]', before any biopsy word appears. Inspecting the organ is the key ingredient; the scope…","st":"tested","i":6,"c":6,"m":[["scope + inspect",10.6],["scope, no inspect",9.0],["colonoscope view",8.5],["colposcopy",8.2],["pipe camera",7.5],["open surgery",4.4],["no organ",2.9],["needle",2.4],["unrelated",-0.1],["difficulty",-4.2]]},{"id":"L32_E35_N42","k":"m","L":32,"E":35,"N":42,"t":"A signed neuron: publishing formulas up, named-entity facts down","l":"Inside the narrow slice of text its expert admits, mostly passages the model has memorized, this neuron takes a sign. It goes positive on the next word of a fixed formula about how material is issued, published or credited, such as wire-service rights notices, the NPR transcript…","st":"tested","i":3,"c":4,"m":[["gov publication",3.2],["tech report",1.9],["named magazine",0.5],["named film",0.4],["press release",0.3],["official pub",0.0],["public document",-0.0],["license (control)",-0.0],["food (control)",-0.2]]},{"id":"L32_E43_N237","k":"m","L":32,"E":43,"N":237,"t":"A neuron for the next item in nursery-rhyme catalogues","l":"This neuron tracks the repeated catalogue frame in familiar children's and folk texts: Dr. Seuss, Rossetti's 'Hurt No Living Thing', Robin Hood ballad lists and Candlemas lore. It fires on the 'nor', 'with a' or 'and' just before the next creature or place is named, and…","st":"tested","i":6,"c":6,"m":[["Seuss",9.4],["Rossetti",4.7],["ballad list",3.0],["nursery list",2.9],["legal halo",0.9],["modern list",-0.0],["code prose",-0.0],["abstract verse",-1.0],["ballad speech",-3.6]]},{"id":"L32_E51_N95","k":"m","L":32,"E":51,"N":95,"t":"Zodiac catalogues versus one person's birth chart","l":"Inside astrological text, this neuron tells apart the system from the person. It fires positive on catalogue slots that list the zodiac's parts: signs, planets, houses, exaltations. It fires about three times more strongly in the negative direction when the text turns to one…","st":"tested","i":6,"c":6,"m":[["own sign",-1.5],["orbital number",-3.0],["astrology word",-4.7],["her chart",-5.8],["sign list",-6.0],["control: kitchen",-6.0],["control: season",-6.0],["zodiac",-7.8],["your chart",-17.1],["wearer",-20.3]]},{"id":"L32_E8_N167","k":"m","L":32,"E":8,"N":167,"t":"A neuron that anticipates re-mentioning a known person","l":"This neuron watches for a person who was named earlier and is about to be named again. It fires positively on the token that opens a credit or identity statement, such as 'About the Author:', a catalog ' /' or an answer turn. It fires negatively on connectives that tie a known…","st":"tested","i":8,"c":7,"m":[["full + initial",5.8],["full + initial",5.7],["full name",4.4],["title + name",3.2],["two people",3.0],["surname only",0.9],["no name",0.5],["organization",0.4],["placeholder",-0.3]]},{"id":"L33_E119_N411","k":"m","L":33,"E":119,"N":411,"t":"A neuron that separates asserted theorem claims from inequality equality cases","l":"In theorem and proof writing, this neuron marks the point where a result has been asserted and its characterizing condition is about to follow. It fires on 'Then ... G is connected' inside a Lemma, up to about +10. The same machinery goes strongly negative, to about −16, when…","st":"tested","i":6,"c":6,"m":[["asserted variant",4.8],["equality case",3.6],["asserted claim",3.3],["Then opener",3.0],["definition halo",0.9],["control prose",0.0],["control math",-0.0],["premise",-0.9],["connective",-3.6]]},{"id":"L33_E128_N160","k":"m","L":33,"E":128,"N":160,"t":"A digit neuron that reads 5 as planets, hormones or taxes","l":"This neuron has two signs. It fires positively on the leading 4 or 5 of numbers in prose about specific exoplanets and numbered asteroids, such as '55 Cancri e' or 'NASA has confirmed 5,600 planets'. The same digits turn it strongly negative in two other contexts: 5alpha-steroid…","st":"tested","i":4,"c":3,"m":[["exoplanet count",6.6],["exoplanet name",6.3],["asteroid",4.7],["geology date",2.7],["beetles (control)",-0.0],["sports (control)",-0.0],["scripture",-4.3],["steroid",-5.7],["tax code",-7.5]]},{"id":"L33_E128_N500","k":"m","L":33,"E":128,"N":500,"t":"Placing genes on chromosome arms versus naming which label is meant","l":"This neuron gives a signed reading of how a passage points at a chromosome or another classified item. Its larger, negative side fires when text places something on a chromosome ('mapped on the short arm'), most strongly in crop QTL papers. There it pushes the model to write…","st":"tested","i":6,"c":7,"m":[["microfilm series",7.1],["trisomy count",4.9],["tax citation",4.8],["blood group",4.6],["topic halo",0.0],["recipe control",-0.0],["generic control",0.0],["crack mimic",0.0],["human locus",-9.1],["plant QTL",-20.0]]},{"id":"L33_E136_N68","k":"m","L":33,"E":136,"N":68,"t":"A neuron that picks a broken copy back up","l":"This neuron tracks copies that have been broken. When the model reproduces an earlier sentence in an altered form (woven through with tags, turned into a URL slug, capitalized, or restarted after a turn boundary), it fires on the token just before the copy resumes. A plain…","st":"tested","i":6,"c":7,"m":[["ALL CAPS",1.1],["NER tags",0.9],["underscores",0.8],["answer restates",0.4],["blank to fill",0.1],["URL slug",-0.0],["no copy",-0.0],["unrelated",-0.0],["parenthetical",-0.3],["verbatim repeat",-0.3]]},{"id":"L33_E142_N398","k":"m","L":33,"E":142,"N":398,"t":"Expecting 'from all over the world', and opposing 'faster than anywhere'","l":"This neuron predicts a distributive worldwide phrase. It fires on the word just before 'from all over the world' or 'around the world' is expected, most strongly on 'from' after tourists or visitors that a famous place attracts, and its activation fades once the phrase is…","st":"tested","i":6,"c":7,"m":[["tourists from",14.0],["artists from",13.0],["visitors from",8.9],["shoppers from",8.5],["enjoyed (no from)",6.7],["sail around",0.6],["control",0.0],["control",-0.0],["superlative",-0.1],["faster than",-2.6]]},{"id":"L33_E149_N42","k":"m","L":33,"E":149,"N":42,"t":"A neuron that bets the next word is 'describe'","l":"This neuron fires on the token that begins a 'describe/depict' word while the word is still unfinished. The clearest trigger is a bare Chinese 描, which reaches about +36 in literary criticism. Finished words like 描述 get roughly half that, and describe stems in other scripts fire…","st":"tested","i":5,"c":9,"m":[["bare 描",32.3],["描述 merged",13.8],["描写 merged",12.9],["描绘 merged",8.6],["sketch",2.4],["scan",1.1],["synonym",0.7],["control",-0.0],["explain",-0.4],["narrate",-0.7]]},{"id":"L33_E170_N110","k":"m","L":33,"E":170,"N":110,"t":"A neuron for needs that only time can meet","l":"This neuron fires on 'need', 'must' or 'give' when what is still required is time: a body healing, a heart getting over a breakup, a team adjusting to a hard language. The same verb asking for an action, an amount, or under negation pushes it negative. Its output promotes the…","st":"tested","i":8,"c":8,"m":[["adapting",8.5],["grief",6.0],["healing",5.4],["courage",2.6],["functional",0.8],["process",0.5],["control",-0.2],["control",-0.2],["compute",-0.9],["negated",-2.5]]},{"id":"L33_E227_N170","k":"m","L":33,"E":227,"N":170,"t":"A recitation-only neuron: positive on restated terms, negative after openers","l":"This neuron sits in a rarely used expert and is silent on almost all text. It speaks only inside passages the model seems to be reciting, such as encyclopedia leads, Lorem ipsum and a few poems. There it turns positive on the pieces of the term being restated and negative on the…","st":"tested","i":3,"c":7,"m":[["mid-sentence restate",0.6],["restated term",0.4],["term first piece",0.4],["single-token term",0.1],["fixed string",0.0],["new term",-0.0],["new name",-0.0],["verse opener",0.0],["control",-0.0]]},{"id":"L33_E73_N511","k":"m","L":33,"E":73,"N":511,"t":"A coordinator that closes named catalogues and cues 'many others'","l":"This neuron reads coordinators and signs what they open. It dips hard at the final 'and / y / и / und' of a catalogue of named things, just before 'and many others', and that dip is what helps the model say 'many'. The effect is several times deeper outside English. It rises at…","st":"tested","i":6,"c":8,"m":[["not only: possession",0.7],["partly…but",0.5],["not only: quality",0.4],["generic list",0.2],["above-all aside",0.1],["plain but",-0.0],["plain and",-0.0],["not only: place",-0.8],["name roster",-5.0]]},{"id":"L34_E137_N330","k":"m","L":34,"E":137,"N":330,"t":"A neuron that remembers where King James put the colon","l":"This neuron decides between a colon and a semicolon at strong breaks in memorized scripture. When the model recites a King James verse it knows by heart, the neuron dives negative over the last few words before the verse's colon and writes that colon; zeroing it lets ';' take…","st":"tested","i":9,"c":9,"m":[["Spanish Bible",2.8],["KJV parody",1.0],["Gettysburg control",0.4],["prose control",0.1],["modern translation",-0.2],["Hamlet",-5.0],["KJV, period next",-5.7],["reworded verse",-6.6],["KJV, colon next",-9.9],["KJV, colon next",-14.3]]},{"id":"L34_E154_N173","k":"m","L":34,"E":154,"N":173,"t":"A slot opens for built objects; a predictable closer shuts it","l":"This neuron in a coding-heavy layer fires when code reaches a place that could hold things it has just built: a fresh container name, an insert call, a list opener, or a ternary choosing between two objects. It goes strongly negative on the last item of an argument list when…","st":"tested","i":5,"c":6,"m":[["new receiver",4.0],["final item",4.0],["insert method",3.9],["short receiver",2.3],["container name",2.0],["list opener",1.3],["reused element",0.8],["statement start",0.6],["log call",0.0]]},{"id":"L34_E171_N403","k":"m","L":34,"E":171,"N":403,"t":"The moment a shy character finally acts, versus holding back tears","l":"This neuron tracks a turning point in stories: is a character about to finally do something they have long wanted to do and held back from? It fires strongly at that moment ('One day, he…') and predicts words like 'gathered' and 'mustered'. It goes strongly negative in the…","st":"tested","i":8,"c":8,"m":[["withheld confrontation",13.5],["long-held love",13.1],["bare turn",10.2],["deep breath",8.3],["hesitation",7.0],["control: weather",5.5],["restraint",5.1],["control: errand",5.0],["composure",-7.6],["held-back tears",-16.2]]},{"id":"L34_E195_N211","k":"m","L":34,"E":195,"N":211,"t":"A neuron that predicts 'characteristics' after machine-performance nouns","l":"This neuron mainly fires negatively. It reads a noun for how a machine, system or material behaves in operation, such as code performance or a car's handling, at the moment 'characteristics' is the likely next word, and it pushes the model toward that word. Other senses of the…","st":"tested","i":5,"c":7,"m":[["owners (control)",-0.0],["names (control)",0.0],["colors (control)",0.0],["problems (halo)",-1.4],["flavor (weak)",-1.9],["handling",-3.6],["torque",-3.6],["noise",-7.8],["speed",-9.7],["performance",-16.4]]},{"id":"L34_E199_N456","k":"m","L":34,"E":199,"N":456,"t":"A signed given-name table read in film credits and births lists","l":"This neuron reads a signed table of given names at the first token of a person's name in film credits, cast lists, births lists and reviews. A few names, such as David, William and Peter, push it positive. Its largest values come after 'Directed by'. A few others, such as…","st":"tested","i":4,"c":8,"m":[["David (core)",8.3],["Peter",3.8],["Anne",3.5],["William",2.2],["John (neutral)",0.6],["'the' (control)",0.0],["Kiki",-0.3],["year (control)",-0.5],["Clarence",-1.8],["Xander",-2.4]]},{"id":"L34_E206_N386","k":"m","L":34,"E":206,"N":386,"t":"A large-ratio neuron whose sign tracks predicate versus fold-multiplier syntax","l":"This neuron marks the moment a large ratio is about to be stated, once the context has shown that one thing far exceeds another. The syntax of the upcoming ratio sets the sign. It is positive on the copula before 'is a thousand times faster', and strongly negative on the article…","st":"tested","i":6,"c":6,"m":[["copula (core)",8.4],["by a factor",8.4],["control",8.4],["control",8.4],["control",6.4],["has + comparative",3.8],["nominal (verb)",3.1],["future tense",2.6],["past tense",1.9],["nominal multiplier",-4.2]]},{"id":"L34_E228_N60","k":"m","L":34,"E":228,"N":60,"t":"An NYC subway-list neuron that turns negative on Regents exams","l":"This neuron tracks open slots in New York text. It fires on the gaps of a subway line list that is still being spelled out (served by the A[,] C) and on the next word of New York program names. It turns strongly negative right where New York students must 'pass the' or prepare…","st":"tested","i":6,"c":6,"m":[["digit list",4.4],["letter list",4.3],["final item",4.3],["name list",2.3],["NY program",2.2],["control",0.1],["control",-0.1],["pass the",-2.0],["prepare for",-2.4]]},{"id":"L34_E229_N52","k":"m","L":34,"E":229,"N":52,"t":"A neuron for what becomes of the dead in Buddhist rites","l":"This neuron sits in an expert that the router sends mostly Buddhist and East Asian religious text. Inside that text it fires on the word just before the story says what happens to the dead or to living beings: freed, saved, reborn. It also fires on the bardo term 中阴 itself. It…","st":"tested","i":6,"c":4,"m":[["hungry spirits",8.8],["unclaimed dead",7.9],["ghosts",7.8],["animals",7.5],["ancestors",6.3],["roof",0.2],["students",0.0],["documents",0.0],["schedule",0.0]]},{"id":"L34_E40_N279","k":"m","L":34,"E":40,"N":279,"t":"Folk-tale traditions versus the word before a famous 'Story'","l":"This neuron tracks two kinds of story talk with opposite signs. It goes positive through scholarly discussion of folk tales as a tradition: fairy-tale and legend collections, the Arabian Nights, and old pamphlet title lists. It goes sharply negative on the single token before…","st":"tested","i":7,"c":6,"m":[["folk genre",3.4],["tale collection",2.7],["collection title",2.3],["antiquarian citation",0.5],["control",-0.4],["other-head title",-0.9],["invented title",-0.9],["known Tale",-7.3],["known Story",-8.5],["known Story",-9.0]]},{"id":"L34_E67_N434","k":"m","L":34,"E":67,"N":434,"t":"A neuron that anticipates 'equal to' in textbook formulas","l":"This neuron fires on the word just before the model expects 'equal to' in a math or physics statement, such as the 'and' in 'the sum of the distances is constant and equal to 2a'. It tracks how confident that expectation is, not the surrounding words: the same 'is constant and'…","st":"tested","i":5,"c":8,"m":[["ellipse foci",26.6],["hyperbola",23.9],["circle radius",22.5],["parabola ratio",18.6],["car speed",17.8],["tree height",17.4],["physics invariant",14.4],["control: color",14.1],["control: mood",14.0]]},{"id":"L34_E74_N208","k":"m","L":34,"E":74,"N":208,"t":"Code line ends: action branch ahead versus value return ahead","l":"This neuron reads line ends in code and forecasts what the function does next. It turns mildly positive after an if, else or except header whose body will print, append or log. It turns sharply negative after the last work statement of a function that is about to return the…","st":"tested","i":6,"c":8,"m":[["st.write",1.7],["append",1.6],["print",1.3],["log",-1.8],["raise",-2.1],["assign (halo)",-2.4],["return",-3.5],["continue",-8.0],["pass",-8.4]]},{"id":"L35_E106_N160","k":"m","L":35,"E":106,"N":160,"t":"A prefix detector whose sign tracks whether the text is English prose","l":"This neuron fires only on a few bound prefixes (non, self, re, semi and their translations), but the sign comes from the language of the surrounding text. In code, Chinese, Russian or German the prefix reads positive. In fluent English prose it reads strongly negative, reaching…","st":"tested","i":6,"c":8,"m":[["dict key",5.0],["code _non",4.7],["Chinese 非",3.4],["German Selbst",2.5],["tech English",2.2],["control",0.0],["control",-0.0],["prose semi",-6.7],["self-help English",-8.5],["legal English",-11.1]]},{"id":"L35_E112_N56","k":"m","L":35,"E":112,"N":56,"t":"A neuron for 'make easy money from home' pitches","l":"This neuron tracks one specific Chinese pitch: easy side income earned online, on a phone or from home. It builds over words like 网赚, 兼职 and 手机赚钱 and peaks on the commas and chat-turn tokens that follow, where it nudges the model to start thinking (<think>) before answering.…","st":"tested","i":6,"c":7,"m":[["game farming",16.4],["phone money",13.9],["online money",10.9],["home gig",9.8],["food delivery",4.0],["street stall",3.5],["hype copy",1.9],["investing",0.8],["learn English",0.8],["weight loss",-0.4]]},{"id":"L35_E126_N254","k":"m","L":35,"E":126,"N":254,"t":"A neuron that anticipates the name of an underground armed group","l":"This neuron fires on the small word just before a text names a clandestine armed group, such as the 'the' in 'assassinated by the' or the comma before the next group in a list of guerrilla movements. It fires before the name, not on it, and silencing it makes names like IRA or…","st":"tested","i":7,"c":7,"m":[["liberation fronts",47.8],["militant groups",40.6],["armed-wing parties",32.2],["one group",23.9],["state forces",16.9],["mafia",15.4],["bands",14.5],["war law",12.9],["civic clubs",12.9]]},{"id":"L35_E136_N17","k":"m","L":35,"E":136,"N":17,"t":"A neuron that signs half-written words: -ity/-ize up, -ate down","l":"Inside one expert of a late layer, this neuron marks unfinished Latinate words and gives them a sign. A word cut off just after the linking -i- of -ity or -ize (utili-, mobili-, stabili-) drives it strongly positive. A stem waiting for -ate or -ation (coordin-, estim-, associ-)…","st":"tested","i":5,"c":8,"m":[["utili (core)",11.4],["mobili",9.9],["stabili",8.7],["no -i- yet",2.0],["whole word",0.5],["-ize halo",-0.1],["adjective",-0.4],["associ (neg)",-7.9],["estim (neg)",-8.4],["coordin (neg)",-12.3]]},{"id":"L35_E13_N263","k":"m","L":35,"E":13,"N":263,"t":"Where did the money go? A neuron that follows a sum","l":"This neuron tracks the fate of a specific sum of money. Once text says £5,000 was raised or earned, it fires hard on the word just before we learn where the money went, what is left, or that none remains. Swap the money for tulips and it goes quiet. Its opposite pole is formal…","st":"tested","i":8,"c":8,"m":[["collected sum",16.2],["raised sum",14.3],["pooled sum",14.0],["no amount",11.3],["signatures",10.4],["hours (halo)",10.3],["formal funds",10.2],["painting",8.5],["tulips",4.7]]},{"id":"L35_E141_N299","k":"m","L":35,"E":141,"N":299,"t":"A neuron that predicts 原: the 'hara' in Japanese place names","l":"This neuron predicts the Japanese place-name piece 原 (hara, 'plain/field'). It fires strongly negative on the token just before it, such as 湿 in 湿原 or the 'ah' in Senjogahara, but only when the context makes 原 likely next. Removing it at 稲 in 稲原 wipes out most of the probability…","st":"tested","i":6,"c":6,"m":[["天津 (positive)",7.7],["control city",0.0],["control digit",-0.1],["草 (草原)",-0.7],["forest halo",-2.5],["河 (河原)",-4.8],["高天 (高天原)",-6.3],["湿 (湿原)",-7.3],["sea halo",-11.5]]},{"id":"L35_E146_N20","k":"m","L":35,"E":146,"N":20,"t":"A Dzogchen neuron that also notices chemists' birth years","l":"This neuron is mainly a detector for Dzogchen, the Tibetan 'Great Perfection'. It goes strongly negative on the word in any language and on the list slot where Dzogchen is about to appear, and that negative value pushes the model to say it. Zeroing it at 'Tantra and' cuts the…","st":"tested","i":7,"c":6,"m":[["math (control)",-0.0],["cooking (control)",-0.2],["mindfulness",-0.7],["Zen",-1.7],["Kashmir Shaivism",-2.9],["Advaita",-5.8],["Mahamudra",-12.3],["Dzogchen",-13.6],["Atiyoga",-16.8],["Great Perfection",-17.0]]},{"id":"L35_E196_N131","k":"m","L":35,"E":196,"N":131,"t":"A have/has neuron: track record ahead versus counted options","l":"This neuron reads the have/has slot and signals what comes next. It goes positive when a track record follows, most strongly when an artist's or writer's work is the subject in a bio ('Her engravings have [been exhibited]'). It goes sharply negative, down to about -20, when a…","st":"tested","i":6,"c":7,"m":[["artwork (core)",5.8],["artwork (core)",5.5],["written work",5.1],["track record",3.5],["impersonal",1.0],["control",0.4],["generic noun",-0.1],["you: options",-4.9],["you: options",-9.2]]},{"id":"L35_E206_N457","k":"m","L":35,"E":206,"N":457,"t":"A neuron that tells founders from successors before either is named","l":"This neuron watches the word just before a person is named and asks what role that person will play. If the coming person is the principal (the founder a brand is named after, the credited author, the figure whose death is the topic), it fires positive. If the coming person is a…","st":"tested","i":7,"c":7,"m":[["remembered",5.5],["credited",5.3],["hailed",3.3],["own role",2.2],["control: much as",-0.0],["control: twice as",0.0],["replacement",-2.4],["successor",-7.2],["heir",-9.5]]},{"id":"L35_E243_N93","k":"m","L":35,"E":243,"N":93,"t":"A math-solution neuron: triangle labels positive, finished answers negative","l":"Inside math solutions, this neuron acts as a sign dial. It turns positive just before the model writes a specific label or count. Its sharpest case is the triangle symbol in one memorized 'Since DF ∥ AC' proof. It turns strongly negative, down to about -12, once the answer is…","st":"tested","i":5,"c":6,"m":[["DF ∥ AC (core)",4.6],["DF ∥ XW",4.1],["DF = 5",3.3],["FD ∥ AC",2.9],["ZY ∥ AC",2.2],["AB ∥ CD",1.8],["PQ ∥ BC",1.1],["non-math control",0.0],["fact control",-0.0]]},{"id":"L35_E32_N14","k":"m","L":35,"E":32,"N":14,"t":"A neuron that anticipates the word biopsy","l":"This neuron, deep in Qwen3.6, fires when a biopsy is about to come next in medical text. It peaks on the token just before the word, or where the word is strongly expected even if never written. It goes mildly negative once a biopsy has already been named and the text moves on…","st":"tested","i":6,"c":8,"m":[["sampling halo",19.6],["neck lump",17.1],["mouth ulcer",16.0],["breast mass",15.8],["changing mole",13.3],["garden control",10.5],["benign cyst",9.8],["tree control",6.7],["biopsy done",6.2]]},{"id":"L35_E41_N139","k":"m","L":35,"E":41,"N":139,"t":"A neuron that predicts заявка, and 'guided by' at its negative pole","l":"This neuron tracks one Russian word: заявка, an application, request or exchange bid. It fires where that word is about to appear. Once the text has made it predictable, the peak moves earlier, onto the verb, a modifier or the first word piece, and it pushes the model toward…","st":"tested","i":5,"c":6,"m":[["trading order",8.4],["заявка noun",6.6],["заявление",4.9],["lawsuit (halo)",4.9],["message",4.6],["' зая' piece",4.0],["hates tariffs,",1.0],["same root",0.2],["control",0.2],["forms tariffs,",-6.8]]},{"id":"L35_E5_N467","k":"m","L":35,"E":5,"N":467,"t":"A neuron that expects the next parallel requirement","l":"This neuron fires on the subject of a new clause that is about to receive the same kind of requirement as the clause before it, as in 'the numerator must be divisible by 6 and the result must be positive', or the third slot in a 'can be any of' counting list. It needs a stated…","st":"tested","i":6,"c":6,"m":[["must + and",6.5],["must + but",5.8],["can + and",3.2],["semicolon",1.9],["period",0.5],["'is' control",0.5],["no requirement",-0.0],["prose control",0.0],["counted set",-0.6]]},{"id":"L35_E85_N124","k":"m","L":35,"E":85,"N":124,"t":"The 'you' in 'create an account if you don't have one'","l":"This neuron fires on the subject of a clause that is about to say the subject doesn't have something, or that something doesn't exist yet, when that thing has already been named. The clearest case is 'Create an account if [you]' in setup instructions, which reaches about +20. It…","st":"tested","i":6,"c":8,"m":[["create account",23.2],["sign up",21.6],["generate key",20.7],["install tool",16.3],["borrow (halo)",15.6],["don't worry",2.8],["existing account",-0.9],["spare key",-1.0],["never-fails",-5.6]]},{"id":"L36_E133_N216","k":"m","L":36,"E":133,"N":216,"t":"A Romance neuron that sets up 'si' after open questions","l":"In Spanish, Portuguese, Italian and French, this neuron fires on the word that opens room for an unsettled issue, such as the 'de' in 'la pregunta de' or the verb 'Verifique'. Its output pushes the next word toward 'si' (whether/if). Removing it at a peak cuts the chance of 'si'…","st":"tested","i":7,"c":8,"m":[["polémica acerca de",30.3],["pregunta de",29.8],["divididas sobre",22.7],["determinar",15.2],["depende de",14.5],["me pregunto",7.2],["libro sobre",5.4],["casa de",1.0],["English",0.0],["¿qué",-7.5]]},{"id":"L36_E143_N316","k":"m","L":36,"E":143,"N":316,"t":"A Dutch neuron that spells out feminine job titles","l":"This neuron predicts feminine job titles in Dutch. When Dutch text, or text about Dutch and Belgian people, is about to spell a woman's role noun such as schrijfster or zangeres, it fires on the stem and pushes the -es and -ster endings. Ablating it drops the probability of the…","st":"tested","i":8,"c":6,"m":[["zangeres",22.3],["danseres",19.9],["schrijfster",19.4],["kunstenares",15.6],["woman's bio",6.7],["places",1.7],["objects",0.9],["man's bio",-1.7],["TV channel",-2.1],["novel title",-2.6]]},{"id":"L36_E147_N203","k":"m","L":36,"E":147,"N":203,"t":"A neuron that predicts 'up' for things brought to the surface","l":"This neuron fires on a verb just before the model expects the particle 'up' in the sense of something coming to the surface: coughing up blood, tearing up old tracks, dredging up the past. It is not a general 'up' detector, because 'she was brought (up)' barely registers. It has…","st":"tested","i":7,"c":7,"m":[["pulled up rails",12.2],["feelings surface",9.4],["spit up",4.4],["cough up",4.2],["dig (halo)",4.1],["coin comes up",2.7],["control",0.3],["not come close",-4.2],["literal pull",-7.2],["pulled off",-14.9]]},{"id":"L36_E163_N177","k":"m","L":36,"E":163,"N":177,"t":"Founders push it up; people being sent or removed push it down","l":"In Russian and German text, this neuron tracks who holds power over whom. It fires positively when a founder, owner or company is the protagonist, most strongly on the 'и' after a title like 'основатель'. It fires negatively, and about twice as hard, when a person is sent…","st":"tested","i":6,"c":6,"m":[["founder",9.5],["incorporator",7.8],["owner",6.3],["CEO",6.2],["creator",5.2],["father (control)",0.6],["employee (control)",0.4],["teacher (control)",-0.3],["arrested",-1.1],["dismissed",-3.8]]},{"id":"L36_E16_N42","k":"m","L":36,"E":16,"N":42,"t":"A Peru-next dial: fires before 'Peru', negative inside Peru","l":"This neuron asks one question at each slot: is the next word 'Peru'? It fires strongly just before Peru appears, especially on the comma after 'Lima', and more when the cue points uniquely to Peru. It answers no with a negative value once Peru is settled and the next slot holds…","st":"tested","i":7,"c":8,"m":[["Lima",25.1],["Lima repeated",25.1],["Arequipa",17.5],["Peru city list",17.5],["Callao",12.7],["Trujillo (ambiguous)",9.2],["Quito (control)",-0.5],["Austin (control)",-1.1],["Madrid (control)",-1.5],["Peru address",-10.5]]},{"id":"L36_E189_N381","k":"m","L":36,"E":189,"N":381,"t":"A neuron that tells viscosity solutions from functional calculus","l":"This neuron tracks which of two areas of mathematics a text is in. It goes positive in viscosity-solution theory for Hamilton-Jacobi equations and peaks on the word just before 'unique viscosity solution'; the founders' names alone, even in a sentence about a bakery, are enough…","st":"tested","i":7,"c":6,"m":[["Isaacs games",36.8],["HJB control",34.6],["founders' names",32.7],["viscosity HJ",31.6],["control: baking",27.1],["heat PDE",23.5],["functional eq.",23.4],["matrix spectral",11.8],["control: poetry",9.6],["functional calculus",-23.7]]},{"id":"L36_E214_N200","k":"m","L":36,"E":214,"N":200,"t":"Where to find it: GOV.UK lights up, BBC iPlayer goes dark","l":"This neuron fires when a text about UK public services such as Universal Credit, visas, the DWP or NHS guidance reaches the point where the official source should come next, and it pushes the model to write 'GOV.UK'. Other countries' government services get weaker versions of…","st":"tested","i":8,"c":6,"m":[["Child Benefit",46.7],["Universal Credit",41.1],["UK visa",30.8],["NHS guidance",22.9],["Australian benefit",13.5],["US benefit",8.5],["recipe (control)",1.8],["album (control)",-0.3],["TV programme",-1.5],["pirate stream",-3.8]]},{"id":"L36_E22_N232","k":"m","L":36,"E":22,"N":232,"t":"Stem of a default placeholder identifier, just before its distinguishing number","l":"L36/E22/N232 fires on the last fixed token of a default or placeholder identifier, at the point where the short distinguishing number comes next. The strong range (+6 and up) comes from a few families: private-range IP prefixes (192.168.[1].3 +13.0, the dot after it +10.5)…","st":"tested","i":6,"c":8,"q":[["p0_01",-3.56],["p0_1",-2.08],["p1",-0.951],["p5",-0.363],["p10",-0.196],["p25",-0.05],["p50",0.05],["p75",0.363],["p90",0.951],["p95",1.45],["p99",2.63],["p99_9",4.91],["p99_99",7.47]]},{"id":"L36_E39_N195","k":"m","L":36,"E":39,"N":195,"t":"Candidates being tested versus options set aside","l":"This neuron tracks whether an item is being tried out or set aside. In math reasoning it fires positive on the candidate the writer is about to test, such as 'the golden ratio' in 'let me try', but stays silent when the same words are stated as fact. It fires about twice as hard…","st":"tested","i":6,"c":6,"m":[["everyday guess",3.1],["digit trial",2.4],["math candidate",1.7],["asserted (control)",1.7],["named guess",1.5],["narration (control)",0.2],["problem (control)",-0.2],["omission",-2.4],["recommendation",-5.3],["'or' alternative",-10.9]]},{"id":"L36_E63_N47","k":"m","L":36,"E":63,"N":47,"t":"A neuron for US party blocs that predicts 'base'","l":"This neuron tracks American party blocs. It fires on the party name itself ('the Republican Party') and on the word just before an expected voting base ('appeal to his'), and it pushes the model to write ' base'. Foreign parties and club 'factions' get only weak responses. Its…","st":"tested","i":6,"c":7,"m":[["possessive",18.1],["ideology adj",17.7],["US party",15.4],["other ideology",10.0],["UK party",8.8],["off-topic",3.3],["fan control",3.0],["politician name",-8.1],["plain name",-9.0]]},{"id":"L36_E83_N429","k":"m","L":36,"E":83,"N":429,"t":"A 'Section' predictor that fires on paper roadmaps but rejects MySQL pointers","l":"This neuron predicts the word 'Section'. It fires hardest at the period that ends a paper's roadmap sentence ('The paper proceeds as follows.'), and more weakly at glosses and back-references in scholarly prose. It turns strongly negative at a narrow set of memorized reference…","st":"tested","i":6,"c":5,"m":[["own system",4.3],["control: hotel",3.6],["own paper",3.4],["control: family",3.2],["SQLite",2.7],["PostgreSQL",1.6],["generic server",1.4],["mysqld",-21.5],["InnoDB",-22.9],["MySQL",-24.5]]},{"id":"L37_E133_N86","k":"m","L":37,"E":133,"N":86,"t":"A team-name finisher that cheers college teams, not junior hockey","l":"This neuron watches for the moment a sports team's name is about to be completed. Its sign depends on which team. For North American college sides it is strongly positive, reaching about +11 at 'Bishop's' before 'Gaiters' in a Canadian university roster. Canadian junior hockey…","st":"tested","i":5,"c":5,"m":[["control: study",7.4],["Minnesota",6.9],["Bishop's (U Sports)",6.4],["Michigan Tech",6.3],["Wisconsin",4.1],["Alabama",1.7],["NFL",0.4],["control: trade",-0.1],["Albion Rovers",-2.2],["WHL junior",-7.2]]},{"id":"L37_E150_N78","k":"m","L":37,"E":150,"N":78,"t":"Before a number: worked out on the page, or reported?","l":"This neuron fires on the word just before a number, and its sign tracks where that number comes from. It goes strongly positive when the figure follows from reasoning on the page: the far end of a 从X到Y range, an '=' result, or a worked 'if a system can process N' example. It…","st":"tested","i":9,"c":6,"m":[["computed range",1.3],["control: own",-0.2],["control: share",-0.4],["fiction",-0.5],["reported, plain",-0.6],["hypothetical",-1.4],["worked example",-2.8],["long inventory",-9.4],["real inventory",-11.8]]},{"id":"L37_E153_N446","k":"m","L":37,"E":153,"N":446,"t":"Before 'single' or before a description: a signed next-phrase forecast","l":"This neuron predicts what kind of phrase comes after a short function word. It goes strongly negative when the next word should be an exactly-one word like 'single' or 'solely' ('there is no', 'cannot be attributed'). It goes positive when an open, describing phrase is coming…","st":"tested","i":5,"c":6,"m":[["signifies a",0.0],["offers several",0.0],["and other",-0.0],["might be a",-0.0],["bare many",-0.0],["control: at",-0.0],["control: the",0.0],["expects a (one)",-0.0],["attributed (solely)",-0.0]]},{"id":"L37_E15_N7","k":"m","L":37,"E":15,"N":7,"t":"A copula neuron that bets on 'heavy' versus a site's condition","l":"This neuron fires at the word right before a predicate, usually 'was' or 'is', and its sign tracks what the coming predicate will describe. If the subject is something being carried or lifted, so weight comes next, it goes strongly negative and pushes the model toward 'heavy'.…","st":"tested","i":6,"c":6,"m":[["harbor",6.2],["stadium",6.1],["tunnel",5.4],["road",4.9],["meeting (control)",2.4],["website (halo)",2.3],["patient (control)",0.3],["feather",-4.1],["boulder",-4.4],["suitcase",-5.4]]},{"id":"L37_E171_N479","k":"m","L":37,"E":171,"N":479,"t":"A neuron that anticipates numbers being printed as hex or binary","l":"This code neuron has two signs. Its large negative side fires when a number is about to be written out as hex, binary or base-36 digits. It reads that expectation from variable names, '#' prefixes and format specs, and it pushes the model toward .toString. Its smaller positive…","st":"tested","i":6,"c":7,"m":[["price (control)",0.5],["label (control)",-3.6],["total (control)",-3.7],["colour code",-13.7],["codec name",-15.3],["hex",-25.9],["bits",-28.3],["binary",-28.9],["base-36",-29.8]]},{"id":"L37_E173_N385","k":"m","L":37,"E":173,"N":385,"t":"A Laravel `$table->unsigned` detector whose negative side tracks sports results","l":"This neuron is a Laravel schema-builder detector. It fires on each `$table->` column line and hardest on a modifier stem like `unsigned` that still needs its type. Removing it shifts the model away from big and unsigned integer columns. Other declarative framework code gives a…","st":"tested","i":5,"c":6,"m":[["unsigned stem",10.4],["big stem",4.9],["full type",4.9],["Django field",3.3],["short receiver",0.3],["chess list",0.0],["Flask config",-0.1],["plain PHP",-0.4],["motocross list",-7.0]]},{"id":"L37_E179_N271","k":"m","L":37,"E":179,"N":271,"t":"A Russian neuron that waits for the genitive to continue","l":"Inside Russian and other Slavic noun phrases, this neuron marks the spot where a genitive is expected or could be extended. It fires on head nouns that still need their 'of-whole' ('нарушить работу…', 'насыщает ткани') and hardest on the 'и' right after a completed genitive…","st":"tested","i":6,"c":7,"m":[["head+genitive",18.2],["body genitive",10.8],["head+genitive 2",9.7],["bare head",8.7],["locative PP",8.7],["instrumental PP",8.7],["English",7.3],["possessor first",4.3],["adjective",-0.9]]},{"id":"L37_E181_N303","k":"m","L":37,"E":181,"N":303,"t":"One neuron reads P as either a letter or a shape","l":"This neuron responds only to the letter P, and its sign reports what P is doing. When P is read as a letter (an alphabet slot, a listed letter, or 'the letter P') it fires strongly positive, up to about +24. When P names a figure in a diagram, as in 'angle P' or 'triangle PQR'…","st":"tested","i":6,"c":6,"m":[["the letter",4.9],["letter list",2.9],["alphabet slot",2.4],["label (control)",2.3],["variable",2.0],["point (control)",-0.7],["region",-3.9],["shape",-7.9],["triangle",-10.1],["angle",-12.2]]},{"id":"L37_E185_N507","k":"m","L":37,"E":185,"N":507,"t":"A Slavic neuron that signals the coming case ending on pronoun stems","l":"Late in the network, this neuron looks at Slavic pronoun and numeral stems whose ending has not been written yet. It then signals which kind of case ending should come next. Frames that call for an oblique case, such as 'є од|ними з' or 'без тво|је', push it strongly positive.…","st":"tested","i":6,"c":4,"m":[["instrumental",7.6],["locative",6.1],["Bulgarian halo",4.2],["dative",1.9],["genitive",1.6],["control: noun",0.4],["control: full word",-0.5],["bare initial",-9.7],["accusative 'на'",-10.2],["accusative",-13.4]]},{"id":"L37_E190_N432","k":"m","L":37,"E":190,"N":432,"t":"A neuron that hears a hedged verdict and writes 'All in all'","l":"This neuron fires at the punctuation that closes a modest, hedged judgment in a review, such as 'a decent little scoop of ice cream.' or 'not a bad showing,'. At that point it pushes the model to write 'All', as in 'All in all'. Understatement matters more than praise or…","st":"tested","i":8,"c":6,"m":[["one complaint",14.7],["hedged complaint",5.7],["hedged praise",5.2],["blunt mediocre",4.7],["litotes",3.2],["superlative",1.2],["modal split",1.2],["neutral control",0.4],["neutral fact",-0.1]]},{"id":"L37_E216_N370","k":"m","L":37,"E":216,"N":370,"t":"A neuron that foresees 'men and women' in Chinese, then damps it","l":"This neuron is a forecast of 'men and women.' In Chinese text it fires at the token just before the model expects 男女, or 男 or 女 as one half of a pair: a comma after 老老少少, 不分 at a bathhouse, a line break in a 男：/女： duet. Its output then pushes that word down. It does not track…","st":"tested","i":7,"c":5,"m":[["changing room",13.1],["bathhouse",10.2],["public bath",9.3],["dormitory",9.0],["dating site",5.9],["bookstore (control)",2.3],["bank (control)",1.5],["restaurant hours",0.0],["machines (control)",-0.0]]},{"id":"L37_E226_N412","k":"m","L":37,"E":226,"N":412,"t":"A neuron that predicts Castilian names, and García as their opposite","l":"This neuron watches for the moment a Spanish name is about to be written, and its sign says which kind. Just before a Castilian town, monastery, chronicler or dictionary is named, it fires strongly positive. Just before 'García', for example on 'Gabriel' or 'Federico', it fires…","st":"tested","i":7,"c":7,"m":[["town list",16.6],["loanword",7.7],["Castilian place",5.8],["Portuguese rival",5.8],["Spanish source",3.0],["garrote",1.5],["control math",-0.0],["control name",-0.2],["before García",-8.1],["before García",-8.7]]},{"id":"L37_E230_N306","k":"m","L":37,"E":230,"N":306,"t":"A math value stated inline versus laid out as listed items","l":"In LaTeX worked solutions, this neuron reports how a math value is presented. It fires positive when a single quantity is stated for the first time in running prose, inside inline math such as \\( V_s = 2 \\) or $t=4$. It fires strongly negative when values are laid out as…","st":"tested","i":4,"c":6,"m":[["inline value",9.6],["inline close",9.6],["inline quad",9.6],["inline variable",3.1],["plain text",0.0],["code",0.0],["inline fraction",-0.6],["inline boxed",-2.1],["boxed tuple",-7.4],["display quad",-8.5]]},{"id":"L37_E50_N101","k":"m","L":37,"E":50,"N":101,"t":"An article neuron: measured results positive, omens and consequences negative","l":"This neuron reads the article ' a' and signs it according to what the coming noun phrase will say. It fires positive when the phrase will characterize a just-reported measurement, spec or data result. It fires negative when the phrase will name what something signifies or…","st":"tested","i":7,"c":7,"m":[["spec figure",7.9],["range idiom",7.0],["measured result",6.1],["figure regarded as",5.8],["static thing",2.1],["control: chat",-0.2],["control: math",-1.6],["sign",-6.0],["as a result",-8.5],["omen",-10.1]]},{"id":"L37_E51_N371","k":"m","L":37,"E":51,"N":371,"t":"A Romance neuron that tracks subject number before the verb","l":"In Spanish, Catalan, Portuguese, Italian and French, this neuron tracks whether the subject of the current clause is singular or plural while its verb is still to come. It goes positive after plural subjects and swings far more strongly negative after singular ones. The…","st":"tested","i":9,"c":8,"m":[["plural",4.3],["plural possessive",2.7],["collective",1.6],["1st plural",0.6],["1st person",0.5],["no subject",-0.5],["coordinated",-2.0],["singular, plural attractor",-3.4],["singular",-8.6],["singular demonstrative",-9.5]]},{"id":"L37_E51_N424","k":"m","L":37,"E":51,"N":424,"t":"A neuron that anticipates Romance 'should', signed by verb form","l":"This neuron predicts a coming 'should' in Spanish, Portuguese, Italian and French instructions. When the next word will be a conjugated obligation verb such as Debes, Deberías or Deve, it fires strongly negative, often on the period or colon just before it, and pushes the model…","st":"tested","i":6,"c":8,"m":[["mid-phrase",-0.0],["plain verb",-9.5],["colon",-19.6],["period→Deberías",-32.2],["¿→Debería",-32.2],["subject→debe",-32.2],["tener que",-32.2],["habría→debido",-32.2],["noun deber",-32.2]]},{"id":"L37_E69_N285","k":"m","L":37,"E":69,"N":285,"t":"A signed word lookup that switches on in name slots","l":"This neuron gives each word a fixed positive or negative value, and that value shows up mainly where a person's or band's name is expected, such as birthday lists, lineups and 'feat.' credits. The router sends tokens to this neuron's expert almost only in those slots. The slot…","st":"tested","i":3,"c":6,"m":[["Jordan",3.4],["Blake (core)",3.2],["non-name word",2.8],["place word",1.3],["neutral name",0.0],["neutral name",-0.1],["Michael",-2.5],["Kristen",-5.0],["Cold",-8.1]]},{"id":"L37_E70_N82","k":"m","L":37,"E":70,"N":82,"t":"A Romance-language neuron that chooses between 'series' and 'seriously'","l":"This neuron answers one question in Spanish, Catalan, French, Portuguese and Italian: is a 'series' word coming, or a 'seriously' phrase? It fires strongly on the slot just before 'sèries/series': workout set counts, a maths or train-model 'la', electrical 'en'. When it fires…","st":"tested","i":6,"c":7,"m":[["rest 'entre'",23.1],["gym count",17.4],["electrical 'en'",14.4],["media list",6.8],["neutral",-0.3],["maths series",-0.7],["shopping count",-0.8],["lookalike idiom",-0.8],["joke pivot",-20.0],["'en serio'",-25.4]]},{"id":"L37_E72_N175","k":"m","L":37,"E":72,"N":175,"t":"A neuron that signs separators: new line versus same line","l":"This neuron reads the first separator after a structured unit and signs it by layout. It fires positive when the unit is finished and the next part starts on a new line, such as the next answer option, an 'Output' header, or <\/think> before a reply. It goes negative when the…","st":"tested","i":5,"c":6,"m":[["inline comma",0.3],["option, newline",0.0],["option (A), newline",-0.0],["option A:, newline",0.0],["sentence end",-0.0],["control: mid-sentence",-0.0],["control: word",-0.0],["display open",-1.2],["inline choice (bold)",-6.9],["inline choice",-8.2]]},{"id":"L38_E127_N74","k":"m","L":38,"E":127,"N":74,"t":"A neuron that suggests 'recover' during errors, then holds it back","l":"This neuron tracks one word, 'recover', and pushes in both directions. In English technical text, once a program has hit an error and the text goes on to say how it is handled, the neuron fires and makes 'recover' more likely next. Once 'recover' or 'recovery' has already been…","st":"tested","i":6,"c":8,"m":[["invalid",5.0],["malformed",4.5],["exception",4.1],["fatal error",3.9],["anomaly (halo)",2.3],["control: normal",0.9],["control: signup",0.6],["control: greeting",0.6],["account recovery",-2.7],["recovery",-5.3]]},{"id":"L38_E13_N249","k":"m","L":38,"E":13,"N":249,"t":"A neuron for deep decimal digits in coordinate lists","l":"This neuron, deep in the network, watches digits. It fires most on the 4th or 5th decimal digit of long numbers stacked in lists such as map coordinates, prices or code arrays. Even there, only two or three of the ten possible digits fire, depending on the digits before them.…","st":"tested","i":4,"c":6,"m":[["core digit 2",37.9],["core digit",27.3],["unselected digit",16.3],["unselected digit 2",16.3],["4th place",16.3],["zero integer part",5.4],["year",3.6],["ISBN group",1.9],["thousands group",1.8],["bare integer",1.1]]},{"id":"L38_E156_N75","k":"m","L":38,"E":156,"N":75,"t":"An Italian-cc versus English-ck spelling dial on c-final fragments","l":"This neuron sits on word fragments that end in c and asks which spelling system comes next. When an Italian word or surname is about to continue with cc or cch (Bian|chi, Cec|chetti), it fires strongly positive. When English spelling expects a k instead (bac|k, qu|ic|kly, 'clic'…","st":"tested","i":6,"c":8,"m":[["Roc(ca)",21.2],["Bian(chi)",20.5],["Cec(chetti)",15.0],["Mac (English)",12.4],["Boc(cherini)",12.2],["Luc(io)",11.1],["Loc(ust)",3.0],["Dec(atur)",1.9],["control Mart",-0.6],["control Ross",-0.8]]},{"id":"L38_E168_N217","k":"m","L":38,"E":168,"N":217,"t":"A neuron for message text inside escaped code strings","l":"This neuron, deep in the model, fires on readable message content inside code strings whose lines are separated by written escapes such as \\n: the words and numbers of a printf message, a file.write line or a JSON body. In the same kind of string it goes negative on layout, such…","st":"tested","i":5,"c":6,"m":[["printf",12.8],["fprintf",6.4],["fputs",4.1],["file write",2.8],["Python print",0.8],["comment",0.3],["sprintf",0.1],["JS error",0.1],["console.log",-0.0]]},{"id":"L38_E179_N404","k":"m","L":38,"E":179,"N":404,"t":"An Arabic 'depression is coming' neuron that also names Redux slices","l":"In Arabic, this neuron anticipates the word depression (اكتئاب). It fires on the token just before that word, for example after 'the doctor diagnosed his condition as', and it pushes the model to write it. Zeroing it roughly halves the model's chance of producing the word. The…","st":"tested","i":6,"c":3,"m":[["sadness, lost interest",12.6],["named depression",9.5],["panic",8.2],["anxiety/insomnia",6.7],["hopelessness",6.0],["knee pain (control)",5.3],["flu (control)",4.8],["travel (control)",2.1],["fairness list",-8.4]]},{"id":"L38_E194_N210","k":"m","L":38,"E":194,"N":210,"t":"A Chinese-fiction neuron: invented names sink it, manner idioms lift it","l":"Inside Chinese fiction, this neuron tracks who is acting and how. It swings strongly negative on the last token of an invented name, such as 方墨[轩] or Zo[ey], that opens a clause as its subject. It swings positive on the middle of a manner idiom that describes how that character…","st":"tested","i":5,"c":4,"m":[["mid-idiom 羞[答]",8.5],["mid-idiom 慢[吞]",8.2],["mid-idiom 意味[深]",6.8],["common noun",-0.3],["pronoun",-4.6],["plain manner",-5.9],["role noun",-5.9],["foreign name",-10.2],["coined name",-13.8]]},{"id":"L38_E216_N365","k":"m","L":38,"E":216,"N":365,"t":"A math neuron for reused names that need a prime","l":"This neuron fires on a single letter that opens an inline math expression at the moment a new object is being named. It fires hardest when the letter reuses the name of an object already on the page, so the new object is a modified copy that must be marked with a prime or…","st":"tested","i":9,"c":7,"m":[["reused name",11.6],["plain letter",7.8],["fresh letter",5.3],["dollar math",3.8],["reused other name",3.2],["glued \\(",2.3],["prose control",0.0],["prose control 2",0.0],["glued paren",-3.9],["glued bracket",-7.3]]},{"id":"L38_E240_N135","k":"m","L":38,"E":240,"N":135,"t":"Correcting survey measurements versus describing planets and stars","l":"Inside Chinese technical text, this neuron acts as a signed topic dial. It fires strongly when measurements of the Earth's surface (leveling, elevations, survey adjustment) must be corrected for error, with peaks near +30. It turns negative when stars, planets or the Earth are…","st":"tested","i":5,"c":6,"m":[["adjustment",28.3],["leveling error",23.2],["elevations",21.8],["Kalman",16.7],["everyday",16.1],["calibration",14.8],["sun",9.0],["planets",5.9],["control: sales",3.6],["control: web",2.3]]},{"id":"L38_E5_N43","k":"m","L":38,"E":5,"N":43,"t":"A neuron that announces 'Coney Island' just before it arrives","l":"This neuron near the top of the model fires on the word just before a place name, and its sign tells you which place is coming. It peaks when the next words should be 'Coney Island', fires more weakly for Brooklyn's other beaches and New York hot-dog talk, and turns negative…","st":"tested","i":8,"c":8,"m":[["Coney cues",22.7],["Coney Island",12.4],["Jones Beach",10.1],["Rockaway",7.0],["Hudson",2.7],["Brighton Beach",2.4],["control: office",0.6],["control: desk",0.3],["Lake Champlain",-3.7],["Peconic River",-7.2]]},{"id":"L38_E70_N338","k":"m","L":38,"E":70,"N":338,"t":"A neuron that predicts the word 'light' in Arabic and Hebrew","l":"This neuron predicts the word for 'light' in the not-heavy sense, Arabic خفيف and Hebrew קל. It peaks at the pivot right after 'not heavy, but' and on the word's first letter, and it extends weakly to easing a burden and to mildness. The same kha letter goes negative when it…","st":"tested","i":8,"c":6,"m":[["heavy, as thought",25.3],["heavy (core)",12.1],["stone-heavy",11.6],["tiring to carry",4.4],["bulky (halo)",3.8],["cheap (control)",1.2],["full→empty",-2.2],["inside→outside",-2.9],["hot (control)",-3.3]]},{"id":"L38_E99_N77","k":"m","L":38,"E":99,"N":77,"t":"A New Jersey context neuron that turns negative on nature writing","l":"This neuron tracks New Jersey context. Once a passage is clearly New Jersey government, catalog or history text, the neuron fires on the small words and place-name pieces that follow, at roughly twice the level that the same formal language gets for other states. Nature and…","st":"tested","i":5,"c":6,"m":[["NJ agency",5.6],["NJ statute",3.7],["NJ history",2.5],["PA deeds",1.3],["VA statute",1.2],["recipes",-0.0],["math",0.0],["MA ecology",-6.4],["Ohio ecology",-7.3]]},{"id":"L39_E137_N162","k":"m","L":39,"E":137,"N":162,"t":"A split first letter: first name or common noun versus place","l":"This neuron fires on a single letter that the tokenizer has split off the start of a word, before the rest of the word appears. Its sign reflects the model's guess about what kind of word is starting. It is positive when context calls for a person's first name or an ordinary…","st":"tested","i":9,"c":6,"m":[["puppy, G",3.6],["called G",2.6],["baby, G",2.0],["town, G",1.6],["group, J",1.2],["baby, M",1.1],["family name",0.4],["whole word",0.0],["no letter",0.0]]},{"id":"L39_E151_N201","k":"m","L":39,"E":151,"N":201,"t":"A 破/沙 versus 平 signal on split Chinese word onsets","l":"Inside one expert that handles the first character of split two-character Chinese words, this neuron reads which character it is. 破 and 沙 push it strongly positive, most of all in internet slang (破防, 沙雕), where it can reach about +14. A few break-and-cut relatives (裂, 割, 杀, 魔)…","st":"tested","i":3,"c":5,"m":[["沙 (core)",10.9],["破 (core)",10.8],["割 (halo)",5.0],["杀 (halo)",4.9],["魔 (halo)",4.8],["裂 (halo)",3.0],["附 (control)",2.7],["恒 (control)",-2.5],["平 (negative)",-3.2],["降 (negative)",-3.6]]},{"id":"L39_E182_N298","k":"m","L":39,"E":182,"N":298,"t":"Romanized Russian state phrases up, an expected Russian down","l":"This neuron has two jobs that share one output, the 'Russ/Ross' name tokens. It fires strongly positive on word endings inside romanized Russian phrases that name a state system or ministry, such as 'Minoborony' or 'iz sostava', in any country's headlines and citations. It fires…","st":"tested","i":7,"c":7,"m":[["Minoborony",13.6],["Crimea sostava",13.3],["tax system v",6.2],["Polish -ko",4.3],["English control",0.2],["Serbian control",-3.1],["football control",-4.5],["Italy sostava",-5.6],["Spanish chemist",-6.6],["German zugunsten",-13.5]]},{"id":"L39_E241_N291","k":"m","L":39,"E":241,"N":291,"t":"A signed detector for unfinished Japanese katakana words","l":"This neuron fires when Japanese text stops partway through a word, usually a katakana loanword split across tokens, such as the イン of インポート in programming help. The fragment sets the sign: イン, グ and モ are positive, while ボ, バ, マ and furigana kana are negative. The word the model…","st":"tested","i":6,"c":8,"m":[["グ",10.6],["イン (core)",9.3],["whole word",9.3],["ド",5.9],["モ",5.7],["ク (weak)",4.3],["English",-0.1],["kanji word",-0.2],["ボ (negative)",-0.5],["マ (negative)",-1.3]]},{"id":"L39_E245_N135","k":"m","L":39,"E":245,"N":135,"t":"Next item: a math function (+) or a physical hazard (−)","l":"This neuron predicts what kind of thing comes next in a list. It turns strongly positive where the next word should be a mathematical form, such as linear, quadratic or exponential, peaking at about +18 on the comma after 'линейная'. It turns negative where the next word should…","st":"tested","i":5,"c":6,"m":[["exponential",9.7],["logistic regr.",4.9],["Newton method",4.9],["linear",4.8],["linear PSU",1.9],["apples",0.3],["wedding dress",-0.1],["moisture, dust",-2.5],["oxidation",-4.2]]},{"id":"L39_E27_N115","k":"m","L":39,"E":27,"N":115,"t":"'Most' in period praise versus 'more' in felt change","l":"This neuron reads the degree words 'most' and 'more' and decides between two moods. It fires when the word opens an appraisal in period prose: a critic praising 'the most disciplined' story, or a 1940s bulletin advising 'to be most effective'. It goes negative when 'more'…","st":"tested","i":6,"c":3,"m":[["period critic",5.2],["committee report",3.2],["control: idiom",-0.0],["control: offer",-0.7],["more than N",-0.8],["scholarly chain",-1.0],["modern likelihood",-2.7],["the more...the more",-3.2],["seemed more",-5.6],["felt more",-7.4]]},{"id":"L39_E44_N152","k":"m","L":39,"E":44,"N":152,"t":"A neuron that anticipates 'time-consuming' and 'energy-hungry'","l":"Late in Qwen3.6, one neuron fires on a resource noun such as time, energy, memory or power. It responds at the moment the model expects that noun to open a cost adjective like time-consuming, energy-intensive or power-hungry, and it peaks in lists of drawbacks. It barely…","st":"tested","i":7,"c":8,"m":[["memory",12.4],["time (core)",11.6],["power",11.1],["odd noun, cost slot",10.4],["halo",6.0],["good-news frame",5.2],["negated energy",4.8],["research referent",0.7],["control",-0.0],["info referent",-1.7]]},{"id":"L39_E53_N496","k":"m","L":39,"E":53,"N":496,"t":"Russian prefixed-word stems up, spelled-out French names down","l":"In Russian text this neuron lights up on the unfinished stem of a familiar prefixed word, such as предполо- in предположим or продол- in продолжайте, and grows stronger as fluent Russian accumulates before it. It goes negative when French names or loans are spelled out syllable…","st":"tested","i":4,"c":5,"m":[["предполож- core",14.5],["продолж- core",13.0],["halo stem",3.1],["weak halo",3.1],["nonce control",1.9],["предупрежд- core",0.6],["German name",0.2],["Latin control",-1.5],["French name",-1.5],["French name",-2.8]]},{"id":"L39_E55_N407","k":"m","L":39,"E":55,"N":407,"t":"A tw-/nin- onset neuron that flips sign on technical labels","l":"This neuron watches for words that begin with a split 'tw' or 'nin' piece. When the piece starts an everyday word or a count, as in 'adolescent twentysomething' or glued spelled-digit strings, it fires strongly. When the same piece, or a digit 2, opens a property label of a…","st":"tested","i":5,"c":6,"m":[["legal count",8.2],["glued digits",8.2],["everyday tw-",7.5],["twist",6.5],["count cue",3.9],["control thr-",0.2],["control whole word",-0.3],["craft property",-3.9],["graph property",-5.1],["digit property",-6.2]]},{"id":"L39_E73_N245","k":"m","L":39,"E":73,"N":245,"t":"A neuron for split-off 战, pushed negative by names and products","l":"This late-layer neuron scores a single character that has been split off as the first piece of a word. Its value adds a character term and a frame term. The character 战 (war or battle) opening an idiom in esports or fantasy battle narration drives it strongest, up to about +9.…","st":"tested","i":4,"c":6,"m":[["战 (core)",9.5],["戰 traditional",7.3],["大 halo",4.4],["戦 Japanese",3.9],["巨 halo",3.3],["斗 war synonym",1.5],["小 negative-leaning",0.4],["兵 war synonym",0.2],["period control",0.0],["English control",0.0]]},{"id":"L39_E74_N15","k":"m","L":39,"E":74,"N":15,"t":"A neuron for numbers expected to start a span","l":"This late-layer neuron fires on a digit that the model expects to start a numeric span. The digit can be the first verse of a scripture citation range, the start year of a reign, or the low end of a measured range, and peaks reach about +10. It reads expectation rather than the…","st":"tested","i":5,"c":5,"m":[["reign year",3.5],["verse range",1.5],["Ada bound",1.1],["verse range 2",1.0],["page number",0.6],["measurement",0.0],["code literal",0.0],["lone number",-0.1],["calendar week",-3.2],["°F conversion",-4.8]]},{"id":"L39_E75_N30","k":"m","L":39,"E":75,"N":30,"t":"A neuron that dips at commas in passages recited from memory","l":"This neuron goes negative on the last word before a comma, but only when the model is reciting a passage it knows by heart and that passage has list-like or parallel structure: scripture, the Constitution, famous problem statements, dataset templates. Ordinary new prose with the…","st":"tested","i":9,"c":6,"m":[["extra word",0.0],["other food",0.0],["control",-0.0],["control",-0.0],["control",-0.0],["other produce",-2.4],["generic noun",-2.5],["close variant",-5.5],["canonical",-9.9]]},{"id":"L39_E81_N104","k":"m","L":39,"E":81,"N":104,"t":"A game-guide neuron: positive at skill names, negative on stats","l":"Inside Chinese video-game guides, this neuron marks the place where a named skill or a hero's capability is about to begin, such as the 冰 that opens an ice spell or 控制和 in 'provides control and protection'. It turns strongly negative on numeric gear-stat lines like '+50%…","st":"tested","i":4,"c":4,"m":[["冰 (ice skill)",7.2],["毒 (poison skill)",5.5],["火 (fire skill)",4.3],["role conjunction",2.4],["闪电 (lightning)",1.4],["noun control",0.4],["place control",-1.0],["stat affix",-2.8],["affix modifier",-3.3]]},{"id":"L39_E91_N438","k":"m","L":39,"E":91,"N":438,"t":"A neuron that votes on 'ed' after p-fragments like bicycle/p","l":"This late-layer neuron fires on the opening fragment of a word and casts a signed vote on whether the next piece is 'ed'. When context predicts a ped- word (bicycle/p-edestrian, back-p-edaling, 7.P-ediatrics), it swings strongly negative, to about -11, and promotes 'ed'. The…","st":"tested","i":4,"c":6,"m":[["Encyclopædia",2.9],["apricots (halo)",1.5],["pepper (control)",0.5],["no p (control)",0.0],["expedition",-0.2],["paediatric",-1.5],["Cupid",-4.0],["pedestrian",-6.9],["back-pedaling",-7.0]]},{"id":"L39_E9_N219","k":"m","L":39,"E":9,"N":219,"t":"A signed dial: 'as' up, 'and' down after a finished clause","l":"This late-layer neuron reads the connective that follows a finished clause in expository prose and gives it a sign. 'As' introducing a reason, result or comparison pushes it strongly positive and nudges the model toward 'as per'. A bare 'and', 'but' or 'yet' continuing a factual…","st":"tested","i":4,"c":8,"m":[["as (core)",3.9],["because",0.8],["since",0.3],["in (control)",0.0],["period (control)",0.0],["while",-1.4],["however",-2.0],["comma and",-3.4],["but",-3.7],["bare and",-3.7]]},{"id":"L3_E110_N317","k":"m","L":3,"E":110,"N":317,"t":"A comma-after-formula detector that only switches on near Chinese","l":"This neuron marks the ASCII comma that comes right after a closed technical item: an inline formula, a backticked identifier, or a group of digits. It fires hardest on '\\),' in Chinese math solutions. The neuron itself computes this signal in any language. However, its expert is…","st":"tested","i":4,"c":8,"m":[["LaTeX comma",2.6],["dollar comma",2.2],["digit group",1.5],["math then period",1.4],["backtick comma",1.1],["bare comma",0.9],["math opener",-0.0],["word comma",-0.0],["opaque fragment",-0.1]]},{"id":"L3_E11_N118","k":"m","L":3,"E":11,"N":118,"t":"A neuron that flags specific odd word joins, not fragments","l":"This early-layer neuron acts as a sparse lookup on the current token together with the piece just before it. Its dominant pole is negative. It fires on the continuation pieces of certain rare multi-piece words, such as brand names like Aimersoft, Xhosa word interiors, OCR-broken…","st":"tested","i":2,"c":4,"m":[["core join",-0.0],["weak join",0.0],["nonce, weak",0.0],["nonce, flat",-0.0],["transliterated",-0.0],["common control",0.0],["control",0.0],["control",-0.0],["Chinese name",-0.8]]},{"id":"L3_E136_N264","k":"m","L":3,"E":136,"N":264,"t":"An early detector for Arabic feminine endings like ta-marbuta","l":"This early-layer neuron marks the moment a split Arabic word closes with an inflectional suffix, above all the feminine ending ة or its informal spelling ه. It reads morphology, not letter shape: the same ات piece fires as a plural ending but not when it belongs to the root.…","st":"tested","i":5,"c":8,"m":[["vocalized ة",5.8],["romanized -rah",3.7],["ه for ة",1.9],["whole-word ة",1.9],["lookalike",1.7],["plural ات",1.2],["root ات",0.0],["English control",0.0],["Dutch name",-0.1],["digit control",-0.1]]},{"id":"L3_E141_N486","k":"m","L":3,"E":141,"N":486,"t":"A neuron that spells out Fabrik, and opposes split 'ta' words","l":"Inside a German-subword expert, this neuron recognizes two specific words from their pieces. It fires positively on 'fab' and 'rik' when they spell Fabrik (factory), in plain use or in compounds like Zinkfabrik. It fires about twice as strongly negative on the word-initial…","st":"tested","i":3,"c":8,"m":[["Fabrik",3.4],["compound Fabrik",3.2],["compound Fabrik",2.9],["French cognate",1.6],["rik relative",0.9],["control",0.0],["control",-0.0],["whole-word Ta",-0.6],["split Ta-",-8.6],["split Ta-",-8.6]]},{"id":"L3_E152_N374","k":"m","L":3,"E":152,"N":374,"t":"A Dutch-only detector for Defensie and Justitie, opposed by berg/borg","l":"Deep in an early layer, one neuron acts as a dictionary entry for two Dutch government words, Defensie and Justitie. It fires on them in any form and stays silent on similar-looking endings in other words. The language check is done before the neuron runs: its expert is selected…","st":"tested","i":4,"c":5,"m":[["Defensie (core)",5.3],["Justitie (core)",5.0],["Kennisnet",2.7],["PSV club",2.3],["Ajax",0.8],["control ministry",0.4],["control ministry",0.4],["control brand",0.1],["-berg name",-1.9],["-borg name",-3.1]]},{"id":"L3_E206_N69","k":"m","L":3,"E":206,"N":69,"t":"A neuron for the 'could' in 'How could you?'","l":"This layer-3 neuron recognizes the moment a disbelieving question locks in: it fires on 'could' or 'can' right after 'How', most strongly after narrative buildup, as in 'How could you?'. It responds more weakly to other question and inversion openers such as 'isn't it' and 'Not…","st":"tested","i":4,"c":6,"m":[["How could",2.0],["How can",1.5],["Isn't it",1.1],["Why should",0.6],["Not only",0.5],["How come",0.3],["How would",0.3],["modal alone",-0.0],["plain noun",-0.0],["split name",-0.7]]},{"id":"L3_E229_N2","k":"m","L":3,"E":229,"N":2,"t":"A signed lookup: ' wisdom' up, Bernard Pur|die down","l":"This layer-3 neuron behaves like a small signed lookup table over token pairs rather than a concept detector. The word ' wisdom' and a few specific names (Silvia, Perez, Ruffner) are positive entries. Surname fragments read together with their own given name, such as Bernard…","st":"tested","i":3,"c":5,"m":[["Ruffner 'uff'",1.7],["Silvia (core)",1.4],["Perez",1.3],["wisdom",1.0],["wrong pairing",0.4],["near-synonym",0.2],["other name",-0.2],["Leonard Moore",-0.6],["Lindy Chamber",-1.0],["Bernard Pur",-1.7]]},{"id":"L3_E245_N250","k":"m","L":3,"E":245,"N":250,"t":"A word lookup table that the router shows mostly in Vietnamese","l":"This neuron works like a small lookup table over specific words rather than a single concept. It goes positive on a few entries, such as Japanese 張 used as a verb, Vietnamese 'đỏ' and 'tiêu', and ' Thom' after René. It goes negative on others, such as 'font', 'tones', and…","st":"tested","i":2,"c":2,"m":[["tiêu (pepper)",0.5],["ớt đỏ (red)",0.5],["dầu hào",0.3],["đường (control)",0.1],["thịt bò",0.0],["muối (control)",0.0],["các lát (control)",0.0],["những miếng",-1.2],["các cung",-1.5]]},{"id":"L3_E249_N395","k":"m","L":3,"E":249,"N":395,"t":"A signed lookup on the first piece of repeated split words","l":"This early-layer neuron gives a sign to the first piece of a word that the tokenizer splits into several pieces. The piece itself mostly decides the sign: ' soc'(a), ' Bre'(kke) and ' Cad'(igan) come out positive, while Latinate stems like ' regul' and ' mon' and some surnames…","st":"tested","i":2,"c":1,"m":[["soc (soca)",0.3],["Bre (Brekke)",0.3],["Cad (Cadigan)",0.3],["Basic Ce",0.3],["the Ce (halo)",0.1],["mon (monomials)",-0.0],["Soft (Softball)",-0.0],["control: in",0.0],["control: digit",0.0],["Esp (Espino)",-0.3]]},{"id":"L3_E32_N32","k":"m","L":3,"E":32,"N":32,"t":"A spelling detector: word-start 'cor' up, 'fic' down","l":"This neuron fires on the spelling piece 'cor' at the start of a word, as in coratge, corrent, corbata or Catalan 'cor' (heart), whatever the word means. It turns strongly negative on the piece 'fic' (ciència ficció, ficar) and mildly negative on 'pos'. The Catalan/Spanish…","st":"tested","i":3,"c":9,"m":[["cor (core)",4.6],["Cor capital",4.1],["heart meaning",1.9],["kor look-alike",1.8],["chor look-alike",0.9],["plain word",0.1],["col control",0.0],["cur control",0.0],["pos (neg)",-1.4],["fic (neg)",-5.7]]},{"id":"L3_E64_N320","k":"m","L":3,"E":64,"N":320,"t":"A neuron that spells out Russian 'ispolz-' from word pieces","l":"This neuron is a spelling detector. It fires when a word beginning with i-s-p-o-l, as in the Russian 'use' stem использ-, is assembled from short pieces and the next piece adds ь, з or z. Capital-letter, hyphenated and Latin-transliterated spellings all reach about +12…","st":"tested","i":4,"c":8,"m":[["caps split",11.7],["other meaning",11.7],["hyphen split",10.0],["Latin translit",8.2],["no И",5.3],["near-miss",2.6],["whole token",2.0],["synonym",0.2],["English control",0.1],["control",-0.0]]},{"id":"L3_E66_N105","k":"m","L":3,"E":66,"N":105,"t":"An abstract-argument neuron: descriptive words positive, belief and realism negative","l":"Inside abstract arguments about truth, meaning or value, this neuron scores the content word that serves as predicate or concept, and the word sets the sign. Descriptive words such as flat, expansive, endowed and economy push it positive. Epistemic and metaphysical words such as…","st":"tested","i":4,"c":7,"m":[["flat (core)",2.8],["expansive",2.2],["economy word",2.1],["conditioned",1.8],["endowed",1.6],["plain word",0.8],["control",-0.0],["control",0.0],["truth word",-0.3],["'the case'",-0.9]]},{"id":"L3_E75_N107","k":"m","L":3,"E":75,"N":107,"t":"A signed word table for the second half of names","l":"This neuron reads the second piece of a two-token unit and looks up a sign for that word. 'Values', 'White' and '.select' come out positive. Surnames like 'Bright' or 'Wilson' come out negative. The word before sets how loud the signal is: a first name like Richard or Bill…","st":"tested","i":3,"c":6,"m":[["Values (core)",0.0],["White",0.0],["Forces (halo)",0.0],["Technical (halo)",0.0],["Wilson (surname)",-0.0],["said (control)",-0.0],["and (control)",-0.0],["paren (control)",0.0],["Na (fragment)",-1.2],["Bright (surname)",-2.2]]},{"id":"L3_E79_N32","k":"m","L":3,"E":79,"N":32,"t":"A neuron for 'Studies' after named fields, plus Rublev","l":"An early-layer neuron in Qwen3.6 responds to the word 'Studies' when it names a field, such as Korean Studies or Zulu Studies. It responds most when the modifier is a people or language name. It also fires on the memorized 'Rub' in 'Andrei Rublev'. A handful of fixed pairs…","st":"tested","i":3,"c":2,"m":[["Rublev piece",1.3],["abstract field",0.6],["Korean Studies",0.0],["Zulu Studies",-0.0],["Latino Studies",-0.0],["lowercase",0.0],["method compound",-0.0],["control",0.0],["control",-0.0],["negative pair",-0.2]]},{"id":"L3_E83_N448","k":"m","L":3,"E":83,"N":448,"t":"A neuron that spells laev-, terg-, dext- and recoils from serv-","l":"This layer-3 neuron is a spelling detector, not a meaning detector. It fires on the wordpiece that completes the Latin roots laev-, terg- and dext- (and weakly sinistr-), whether the word means 'left', 'smooth', 'detergent' or is made up. Its opposite pole is the serv- piece, as…","st":"tested","i":3,"c":8,"m":[["laev (core)",4.3],["terg (core)",3.2],["dexter (core)",2.2],["sinister (weak)",1.2],["vestib (halo)",0.5],["nerv (control)",0.0],["cerv (control)",0.0],["clever (control)",0.0],["Serv (negative)",-2.0],["serv (negative)",-2.7]]},{"id":"L3_E9_N215","k":"m","L":3,"E":9,"N":215,"t":"A neuron for the suffix that makes 'parallel' into 'parallelize'","l":"This neuron is a narrow word-part detector. It responds to the -ize ending, and its counterparts in other languages, almost only when that ending turns 'parallel' into a verb or noun: parallelize, paralelización, 并行化. Close relatives such as perpendicularize get partway. Other…","st":"tested","i":4,"c":8,"m":[["core",4.3],["prefixed core",4.3],["UK spelling",3.8],["relative stem",3.5],["concurrency -ize",1.8],["ordering -ize",1.6],["generic -ize",-0.0],["look-alike",0.0],["no suffix",0.0],["control",0.0]]},{"id":"L4_E114_N417","k":"m","L":4,"E":114,"N":417,"t":"A company's first move: opening business verbs versus record nouns","l":"This neuron marks the moment a newly introduced business first does something. It fires on the opening verb, as in 'A bank is offering' or 'A company is announcing', most strongly at the start of a problem or item. Person subjects, document subjects and non-progressive verbs…","st":"tested","i":5,"c":6,"m":[["announcing",3.8],["offering (core)",3.2],["considering",1.9],["non-progressive",1.6],["giving",1.5],["state adjective",-0.1],["adjective control",-0.1],["movement (negative)",-2.2],["activity (negative)",-3.0]]},{"id":"L4_E126_N198","k":"m","L":4,"E":126,"N":198,"t":"A Spanish bean-word neuron with a faint division-slash side","l":"This neuron is mainly a detector for Spain-Spanish bean words. Judías, alubias, fabes and judiones drive it strongly negative on every piece of the word, especially after 'las' in a recipe. Other legumes like lentejas and frijoles barely register. A much weaker positive side…","st":"tested","i":4,"c":8,"m":[["cebollas (control)",0.3],["frijoles (control)",-0.3],["lentejas (control)",-0.4],["croquetas (halo)",-0.8],["sandías (neighbor)",-2.7],["habichuelas",-2.8],["alubias",-3.8],["judiones",-5.7],["judías",-5.8],["fabes",-8.0]]},{"id":"L4_E13_N231","k":"m","L":4,"E":13,"N":231,"t":"A neuron for 'Lowell Observatory', with name-specific signs elsewhere","l":"This neuron is a near-private detector for the word 'Observatory' right after 'Lowell'. The router hands the token after 'Lowell' to this expert as its top choice, and the neuron then picks Observatory over other head words. Elsewhere it gives weak positives on the ends of place…","st":"tested","i":3,"c":6,"m":[["core",4.0],["lowercase",2.2],["misspelled",1.7],["other head",1.6],["place ending",1.6],["weak head",0.2],["other observatory",0.0],["control",0.0],["negative place",-1.0],["negative name",-1.9]]},{"id":"L4_E148_N348","k":"m","L":4,"E":148,"N":348,"t":"A neuron for roles typed by their filler: 'their animal hosts'","l":"This neuron responds to a role noun, such as host, substrate, companion or allies, when the word before it says what kind of thing fills the role. It responds most when a possessor in context holds that role, as in 'their animal hosts' or 'its protein substrate', and reaches…","st":"tested","i":6,"c":8,"m":[["protein substrate",6.7],["dragon companion",6.7],["bacterial hosts",5.7],["robot chefs",5.6],["animal hosts",5.5],["Islamist allies",3.4],["travel companion",0.1],["animal cells",0.1],["one-parent families",-0.1],["faculty staff",-0.3]]},{"id":"L4_E162_N342","k":"m","L":4,"E":162,"N":342,"t":"A multilingual short-circuit detector, kept electrical by its router","l":"This neuron detects the word 'short circuit' at the moment the text commits to it, in English, German, Russian, Chinese and especially Spanish, where 'cortoc' fires before the word is finished. The neuron itself barely cares whether the use is electrical or figurative. In…","st":"tested","i":6,"c":8,"m":[["cortocircuito",5.5],["short-circuit",5.4],["short circuit",5.2],["Kurzschluss",4.8],["dead short",4.2],["ground fault",1.0],["open circuit",0.9],["control: delay",0.0],["control: noise",-0.1],["blocking",-1.3]]},{"id":"L4_E170_N35","k":"m","L":4,"E":170,"N":35,"t":"A multilingual promise reader that fires on non-English pledge words","l":"This neuron goes strongly negative when a word names a promise or pledge and is used to mean an actual undertaking, in any language: Indonesian berjanji, Bengali প্রতিশ্রুতি, Japanese 約束, German Zusage. The word alone and the act of pledging alone each give part of the effect…","st":"tested","i":6,"c":8,"m":[["Hebrew halo",0.2],["hope",0.1],["eat (control)",0.1],["guarantee",0.0],["go (control)",-0.0],["pledge",-1.7],["commit",-1.8],["swear",-2.9],["promise",-9.6]]},{"id":"L4_E190_N80","k":"m","L":4,"E":190,"N":80,"t":"A neuron for years that name an edition, not a date","l":"This layer-4 neuron marks a year that names one edition of something: the decade digit in 'India Year Book 2017' or 'General Election 2019'. The same year used as a date ('published in 2017') leaves it nearly silent. The router sends year digits to this expert regardless of…","st":"tested","i":5,"c":8,"m":[["yearbook",6.3],["observance day",6.0],["election",5.7],["Olympics",5.3],["annual report",4.3],["data label",3.8],["product year",2.0],["season",0.7],["fiscal year",0.6],["plain date",0.1]]},{"id":"L4_E237_N357","k":"m","L":4,"E":237,"N":357,"t":"An 'early'/早 detector that only works in official prose","l":"This neuron responds to a single word family, English 'early' and 'earliest', Chinese 早 and 尽早, and Vietnamese sớm, but only when its expert is chosen, which happens in formal policy or government prose. Close synonyms like 'timely' or 'prompt' barely register. A few specific…","st":"tested","i":4,"c":5,"m":[["earliest",1.2],["calendar early",0.9],["early",0.5],["ASAP halo",0.4],["prompt (control)",0.1],["policy word",0.0],["timely (control)",-0.1],["cattle (weak)",-0.1],["flock",-0.8],["poultry",-1.6]]},{"id":"L4_E23_N26","k":"m","L":4,"E":23,"N":26,"t":"A neuron that finishes Anth- names: Anthozoa, Anthrozoology, Anthoupoli","l":"This neuron completes rare Greek names that begin with Anth-. It fires on the oz, zo or ou piece that follows the stem, as in Anth|oz|oa, Anthro|zo|ology and Anth|ou|poli. Two filters combine. The neuron itself prefers that piece after a th- stem, and the router admits the token…","st":"tested","i":3,"c":8,"m":[["Anth|oz (core)",7.4],["Myth|oz",4.9],["Xanth|oz",4.3],["Ant|oz",3.3],["Acanth|oz",2.3],["Anth|os",0.8],["Anth|rop",0.0],["Mes|oz control",-0.0],["Am|az control",0.0],["rare name",-1.0]]},{"id":"L4_E250_N168","k":"m","L":4,"E":250,"N":168,"t":"A pair-lookup neuron: 'Schedule A' up, 'Six Sigma Green' down","l":"This layer-4 neuron behaves like a small lookup table of memorized token pairs, not a concept. Its strongest entry is a capital letter right after 'Schedule' (about +0.9). A few unrelated stock phrases, such as ATTACH|MENT, F|ife and disclaimer 'treatment', reach similar…","st":"tested","i":3,"c":4,"m":[["Schedule (core)",0.8],["Class",0.2],["Coupon",0.1],["all caps",-0.0],["lowercase",-0.0],["Form",-0.0],["Exhibit",0.0],["Grade (control)",0.0],["Book (control)",-0.0]]},{"id":"L4_E255_N249","k":"m","L":4,"E":255,"N":249,"t":"A neuron that completes familiar name pairs, led by 'American fiction'","l":"This neuron responds to the second word of familiar two-word names and phrases. It responds only when the whole pair is known: 'Harlem Nights' fires, but 'Boogie Nights' barely registers. Each second word carries its own direction. 'fiction', 'Front' and 'Nights' go positive…","st":"tested","i":3,"c":7,"m":[["core",3.7],["novels variant",1.7],["familiar title",1.0],["familiar name",1.0],["halo genre",0.6],["control: other pair",0.0],["control: other pair",0.0],["control: no name",-0.0],["negative name",-0.9],["negative title",-2.3]]},{"id":"L4_E255_N29","k":"m","L":4,"E":255,"N":29,"t":"A neuron that completes 'Carol Ann' to Carol Ann Duffy","l":"This neuron is a one-bigram detector for the name 'Carol Ann'. The router sends the token right after the given name Carol to this expert. There the neuron fires hard and negative only when that token is ' Ann', near -12, and this write is what lets the model predict ' Duffy'.…","st":"tested","i":6,"c":8,"m":[["halo: Pak",0.8],["halo: Kay",0.7],["Ann, no Carol",0.0],["control",0.0],["Carol + other",-1.7],["Carol Anne",-4.2],["Carole Ann",-6.4],["Karol Ann",-6.5],["Carol Ann",-11.5],["lowercase carol",-11.5]]},{"id":"L4_E28_N492","k":"m","L":4,"E":28,"N":492,"t":"A neuron that spells out Python plotting library names","l":"This neuron recognizes the names of Python plotting and notebook tools. It fires on the inner piece of 'Altair' and, more weakly, on pieces of Bokeh, Streamlit, Matplotlib and IPython, and on the '%' that opens a Jupyter magic line. The neuron itself responds to spelling, not…","st":"tested","i":4,"c":6,"m":[["lowercase altair",5.4],["Altair (core)",4.7],["Bokeh",1.9],["Matplotlib",1.6],["Streamlit",1.5],["netCDF (halo)",0.6],["Excel (control)",-0.0],["pandas (control)",-0.1],["pfp (negative)",-0.2],["_GV_ (negative)",-0.4]]},{"id":"L4_E50_N361","k":"m","L":4,"E":50,"N":361,"t":"Drew Miller up, Drew Shirley down: a sparse word-pair lookup","l":"This neuron acts like a tiny lookup table over word pairs. A cue token, usually the first name Drew, sends the next token to the neuron's expert, and the identity of that next token sets the sign. Drew Miller is strongly positive, Drew Shirley is strongly negative, and Drew…","st":"tested","i":3,"c":5,"m":[["core",0.9],["core",0.8],["core",0.6],["quoted cue",0.3],["common surname",0.0],["no Drew",0.0],["not routed",0.0],["mild negative",-0.4],["mild negative",-0.5],["negative pole",-1.1]]},{"id":"L4_E59_N213","k":"m","L":4,"E":59,"N":213,"t":"A 'frequency' and ω detector that only listens inside quantum physics","l":"This neuron is a word detector for 'frequency'. It fires on the word itself, its translations, and the symbol ω, but only when its expert is selected, which happens in quantum and condensed-matter physics text. In that text, 'phonons at lower frequency' fires; 'a guitar string…","st":"tested","i":5,"c":7,"m":[["frequency",4.0],["plural",3.6],["abbreviation",3.2],["ω symbol",1.9],["halo: energy",0.2],["halo: wavelength",0.2],["control",-0.2],["weak negative",-0.4],["control",-0.4],["negative pole",-1.5]]},{"id":"L4_E87_N144","k":"m","L":4,"E":87,"N":144,"t":"After a name stem: News, Book, Social up; name fragments down","l":"This early neuron reads the word that directly follows a short name or brand stem. The router picks out that slot, the gate opens across it, and the word itself sets the sign. 'News' in outlet names such as ACM Tech News pushes it highest, followed by Book, Social and Day.…","st":"tested","i":3,"c":6,"m":[["News (core)",1.8],["glued News",1.3],["Book",0.7],["Review (halo)",0.5],["Social",0.4],["outlet synonym",0.0],["control: and",-0.0],["control: period",0.0],["City",-0.4],["name fragment",-1.4]]},{"id":"L5_E113_N201","k":"m","L":5,"E":113,"N":201,"t":"A neuron for Shun, strongest beside Yao and Yu","l":"This neuron is a detector for the name Shun. It responds to the 'un' of Shun in almost any proper name: Japanese names, a Qing reign title, a pet. It fires hardest, up to about +9, when Shun is the legendary sage emperor listed beside Yao and Yu, and in that setting it also…","st":"tested","i":4,"c":6,"m":[["Shun (sage)",6.3],["Shunzhi",6.3],["Debora (halo)",3.0],["Shang",2.2],["Yu",1.7],["Seker (halo)",1.2],["Wang (control)",0.3],["Zhao (neg)",0.2],["Tom (control)",0.2],["Zhang (neg)",0.1]]},{"id":"L5_E131_N350","k":"m","L":5,"E":131,"N":350,"t":"A signed word-value table read on the second word of a term","l":"This neuron reads the second word of a short two-word term, such as a modifier plus its head or the word after 'X and'. A gate that opens after a modifier, multiplied by a signed value that belongs mostly to the word itself, gives the output. Health and Environmental peaks near…","st":"tested","i":3,"c":7,"m":[["core head",0.6],["positive word",0.0],["mild head",-0.0],["modifier, weak neg",-0.0],["negative word",-0.0],["control: function",0.0],["control: verb",-0.0],["control: digits",-0.1],["stacked modifier",-0.9],["brand, negative",-1.1]]},{"id":"L5_E144_N198","k":"m","L":5,"E":144,"N":198,"t":"A neuron for the words that name what you can control","l":"This neuron fires on the word that closes a phrase marking out what someone can affect, such as 'focus on what you can control', 'it is within your power' or 'each of us can do our part', and it does so across languages. Plain claims of ability or empowering pep talk stay weak.…","st":"tested","i":5,"c":6,"m":[["control (core)",4.4],["power",4.1],["influence",3.6],["reach (halo)",2.5],["rights",1.2],["means",0.9],["home (control)",0.9],["budget (control)",0.5],["borders (control)",0.4]]},{"id":"L5_E15_N15","k":"m","L":5,"E":15,"N":15,"t":"A detector for 'not rated' in rating fields","l":"This neuron is a narrow word detector. It responds to the English word 'rated' when something marks it as absent: 'not rated', 'never rated', 'Unrated'. It is strongest in rating fields and film labels, where it reaches about 4. Plain 'rated', other negated review words…","st":"tested","i":3,"c":8,"m":[["not rated",4.2],["never rated",3.6],["not yet rated",3.4],["Unrated",3.1],["not reviewed",2.2],["not ranked",1.7],["bare rated",1.4],["highly rated",0.7],["N/A control",0.3],["pending control",0.2]]},{"id":"L5_E163_N41","k":"m","L":5,"E":163,"N":41,"t":"A Meta detector that also hears Facebook and Zuckerberg","l":"This neuron fires on the capitalized word Meta in every sense: the company, MetaQuotes, a woman named Meta, or an HTML tag. Look-alikes such as Metro and Metal stay silent. Meta-linked words with no shared letters, such as Facebook and Zuckerberg, reach about two-thirds of that…","st":"tested","i":4,"c":7,"m":[["core",7.4],["name piece",7.4],["ticker caps",4.8],["Facebook tier",3.7],["founder",2.9],["other platform",1.5],["control: big tech",0.3],["control: look-alike",-0.0],["control: look-alike",-0.1],["name suffix",-1.0]]},{"id":"L5_E199_N455","k":"m","L":5,"E":199,"N":455,"t":"A neuron that flags repeated symbols across a chemical bond","l":"This layer-5 neuron watches what follows a dash, slash or equals sign in chemistry-style text. If the second symbol repeats the first, as in Cl-Cl or Br-Br, it swings strongly negative, reaching about -9 in bond-energy tables. If a different symbol follows, as in C–S or Li–S, it…","st":"tested","i":7,"c":8,"m":[["C–S (core)",2.2],["bond word (halo)",1.0],["C–N",0.7],["C–O",0.5],["C–B",0.4],["C–C (repeat)",0.3],["control",0.2],["control",0.0],["C–H (H bias)",-0.3],["C–Cl (halogen)",-0.4]]},{"id":"L5_E210_N308","k":"m","L":5,"E":210,"N":308,"t":"A neuron for compressor names like ZFP, SZ and Blosc","l":"This neuron recognizes the names of data compressors, above all the scientific array compressors ZFP, SZ, MGARD and Blosc, which push it up to about +10. A second, weaker trigger is the word that marks the compression type, 'lossy' or 有损. General codecs such as gzip or JPEG…","st":"tested","i":5,"c":7,"m":[["ZFP",8.7],["SZ",7.0],["Blosc",6.5],["lossy",5.6],["MGARD",5.4],["LZ4 (general)",3.5],["gzip (general)",2.9],["control: agency",0.8],["control: variable",0.6],["technical cost",-1.8]]},{"id":"L5_E24_N403","k":"m","L":5,"E":24,"N":403,"t":"A neuron that separates boys and girls from someone's parents","l":"This neuron responds to plural nouns that name people by sex or age, such as boys, girls, men and women. It fires most when a stock adjective comes first, as in 'little boys'. It flips sign for people framed as someone's family: words for parents in any language, and households…","st":"tested","i":5,"c":8,"m":[["little boys",2.0],["young girls",1.2],["old men",0.7],["bare women",0.7],["singular",0.0],["object control",0.0],["children",-0.2],["surname family",-1.2],["grandparents",-1.5],["foster parents",-1.5]]},{"id":"L5_E253_N22","k":"m","L":5,"E":253,"N":22,"t":"A phrase-closer neuron: familiar completions up, title-ending verbs down","l":"This neuron reads the word that closes a short phrase. It turns positive when a familiar frame is completed by its expected word: 'forever' after 'couldn't last' or 'wait here', 'Jesus' in 'Who Was ___', 'sentence' in 'Given the ___'. It turns negative when a quoted title ends…","st":"tested","i":3,"c":3,"m":[["Jesus (core)",1.7],["History of Beer",1.0],["Built Cars",0.6],["Napoleon (halo)",0.2],["Moses (halo)",0.2],["Sold Cars (ctrl)",0.1],["Rain (ctrl)",0.0],["Cars I Built",-0.6],["Jack Built",-0.7],["Love Go",-0.8]]},{"id":"L5_E255_N21","k":"m","L":5,"E":255,"N":21,"t":"A first-mention detector for the name Terri","l":"This neuron is a spelling detector for the name Terri. It fires on the 'ri' piece only when 'Ter' comes directly before it, and fires hardest at the name's first mention, reaching about 10 in author lists. Near-miss names like Kerri or Teri fire weakly. Repeats, or a word like…","st":"tested","i":4,"c":8,"m":[["Terri",9.8],["lowercase",8.1],["Cyrillic",2.6],["Kerri",1.6],["Teri",1.0],["suppressed",0.9],["Sherri",0.9],["Terry (control)",0.0],["Bob (control)",0.0]]},{"id":"L5_E29_N214","k":"m","L":5,"E":29,"N":214,"t":"A signed lookup table for pieces of rare names","l":"This layer-5 expert neuron acts as a small signed lookup table for pieces of rare names. On certain two-token spellings, such as Far+qu or Michael+Naz, it fires positive. On others, such as Abar+is, Zermelo–Fra or Last+Kiss, it fires strongly negative, reaching about -1.…","st":"tested","i":3,"c":7,"m":[["X. Bush",0.8],["Farqu (core)",0.6],["Naz (core)",0.5],["Clark (weak)",0.4],["Elton",0.3],["Abaris (neg)",-0.0],["G. Bush (ctrl)",-0.0],["Iris (ctrl)",0.0],["Farrokhz (neg)",-0.6],["Dezhao (neg)",-0.7]]},{"id":"L5_E64_N208","k":"m","L":5,"E":64,"N":208,"t":"A closed-gate neuron that spells out 'bened' at its second piece","l":"This neuron responds to the spelled-out word stem 'bened' at its second piece, whether it appears in benediction, Benedict's reagent, benedictina or a Latin species name. Meaning does not matter. A few other stems such as CATAL and Priest reach the same height. Interior…","st":"tested","i":2,"c":1,"m":[["bened (core)",0.5],["Bened",0.4],["Serp (halo)",0.4],["Vened (partial)",0.2],["CATAL stem",0.2],["gened (control)",0.1],["whole word",0.0],["Hamb (neg)",-0.2],["nigr (neg)",-0.3]]},{"id":"L5_E68_N315","k":"m","L":5,"E":68,"N":315,"t":"A signed lookup for specific word pairs: Karen Kot versus night side","l":"This neuron responds to a small, fixed set of specific two-word pairs, firing on the second word. The pair determines the sign. Some surnames after given names and some list compounds (Karen Kot, walking football) push it up. Sleep and day-night compounds (night side, sleep…","st":"tested","i":3,"c":4,"m":[["name pair",0.7],["list compound",0.3],["core compound",-0.0],["name variant",0.0],["weak neighbour",0.0],["control",0.0],["control",-0.0],["sleep pair",0.0],["active cultures",-0.3],["day-night pair",-0.4]]},{"id":"L5_E74_N240","k":"m","L":5,"E":74,"N":240,"t":"A neuron that remembers the Dutch verb trekken","l":"This neuron is a memorized detector for the Dutch verb trekken (to pull or move). It fires on the token that completes trek or trok, peaking above +6, and stays active for a word or two afterward. A second, unrelated trigger is ' Don' in Dutch surnames like van de Donk.…","st":"tested","i":3,"c":4,"m":[["trek (core)",5.9],["trek zich",5.9],["trok (past)",4.3],["treur (lookalike)",2.1],["duw (push)",0.1],["loop (control)",-0.0],["tref (lookalike)",-0.1],["zie (control)",-0.1],["verliefd (neg)",-2.4]]},{"id":"L5_E92_N391","k":"m","L":5,"E":92,"N":391,"t":"An 'absorb' detector that fires negative, with a harmless positive leak","l":"This neuron is a detector for the word 'absorb' in English and in languages with their own absorb stem, and it signals by going strongly negative. The negative lingers over the next few words. Synonyms like 'soak up' or 'take in' produce weaker dips. Its only positive responses…","st":"tested","i":4,"c":9,"m":[["accept (halo)",0.3],["release (control)",0.1],["protect (control)",0.1],["digest (control)",-0.0],["take up",-0.4],["take in",-1.2],["soak up",-2.1],["reabsorb",-2.5],["absorb",-2.9],["adsorb",-3.0]]},{"id":"L6_E103_N423","k":"m","L":6,"E":103,"N":423,"t":"A neuron that spells Messalina, and weakly her lookalikes","l":"This neuron is a spelling detector for one name: Messalina, the Roman empress. It fires on the name's final piece, around +9, whatever the language, case or surrounding text. Close respellings like Massalina or Messolina fire partway, scaled by how near they are to the real…","st":"tested","i":3,"c":8,"m":[["Messalina",10.2],["one-s spelling",9.5],["ending swap",7.3],["stem swap",5.1],["Cass- stem",3.5],["short Mess-",1.1],["other -ina name",0.3],["ending alone",0.0],["split onset",0.0],["control",-0.1]]},{"id":"L6_E162_N509","k":"m","L":6,"E":162,"N":509,"t":"A neuron that completes calendar dates, and dislikes dated book titles","l":"This neuron marks the moment a calendar date is completed: it fires on the decade digit of a year that follows a month and day, such as the '1' in 'May 3, 2016', and fires less for month-only dates or unexpected eras. A year by itself does almost nothing. When a year only dates…","st":"tested","i":5,"c":8,"m":[["full date",1.8],["month-year",1.5],["holiday-year",1.2],["day-first date",1.1],["older decade",0.9],["bare year",0.4],["bare-year comma",0.4],["dated title",0.3],["no date",-0.1],["no year",-0.3]]},{"id":"L6_E193_N70","k":"m","L":6,"E":193,"N":70,"t":"Finishing words in familiar frames, against clock digits after dates","l":"This neuron fires on the word that finishes a short predicate inside a frame the model recognizes: 远 'far' at the end of a dusk line in Chinese verse, 'slowly' after 'die', 'alone' after a worker in rate problems, and 'young' in story openings. Its expert is selected only inside…","st":"tested","i":4,"c":6,"m":[["远 far (core)",1.3],["静 quiet",0.6],["绿 green",0.6],["暗 dark",0.6],["高 tall (halo)",0.5],["白 white",0.4],["人 person (control)",0.4],["近 near",0.3],["digits (control)",0.0],["的 (control)",-0.1]]},{"id":"L6_E196_N381","k":"m","L":6,"E":196,"N":381,"t":"A register switch: claims in prose versus math work in progress","l":"This layer-6 neuron tracks register at clause seams. It reads positive where non-math expository prose presents a claim to a reader, strongest when an expert 'now shows that'. It reads negative inside math working, deepest at 'Let me' right after a constraint like 'must be a…","st":"tested","i":5,"c":6,"m":[["expert presenting",2.4],["recipe step",-0.1],["detective clue",-0.1],["poem line",-0.2],["password rule",-0.4],["derivative result",-0.9],["math doubt",-1.1],["gcd constraint",-1.2],["perfect square",-1.3]]},{"id":"L6_E21_N463","k":"m","L":6,"E":21,"N":463,"t":"A neuron that looks up 'Duda', especially in pattern-recognition references","l":"This neuron is essentially a lookup for one string: the name Duda, best known from the classic Pattern Classification textbook by Duda, Hart and Stork. It fires on 'Duda' in any sense, and a technical or bibliographic setting roughly doubles the response. Co-authors Hart and…","st":"tested","i":3,"c":6,"m":[["Duda",19.1],["Andrzej Duda",12.0],["Sedgewick",2.4],["Gonzalez",1.8],["Buda (control)",1.6],["Hart",1.4],["Smith",-0.0],["no name",-0.0],["James R. Brown",-1.4]]},{"id":"L6_E220_N188","k":"m","L":6,"E":220,"N":188,"t":"A rivet-and-hole neuron that the router limits to workshop text","l":"This neuron tracks a short list of words for holes and joining parts, such as fasteners, rivets, holes and drilling, but only in workshop and engineering text. That domain filter comes mostly from the router: on a bare everyday 'hole' the neuron would respond, but its expert is…","st":"tested","i":4,"c":8,"m":[["fastener (core)",4.1],["hole",3.8],["rivet (core)",3.6],["bolt",1.7],["slot (halo)",1.2],["bore",0.4],["screw (control)",0.2],["weld (control)",0.2],["frame word",-0.1],["name (neg.)",-1.1]]},{"id":"L6_E35_N283","k":"m","L":6,"E":35,"N":283,"t":"A detector for the symbol k, and above all k_0","l":"This neuron is mainly a detector for the letter k used as a math or science symbol. It fires most strongly, and negatively, on the index right after k_, as in k_0 (down to about −9). The bare k in k-means or 'rate constant k' fires more weakly. Other letters in the same slot…","st":"tested","i":3,"c":6,"m":[["letter m",0.0],["control a",-0.0],["control x",-0.4],["word control",-0.4],["letter r",-1.5],["letter j",-1.8],["capital K",-3.9],["kappa",-4.9],["k (core)",-9.0]]},{"id":"L6_E42_N316","k":"m","L":6,"E":42,"N":316,"t":"A neuron for 'you get the idea': leaving the rest unsaid","l":"This neuron responds to one English idiom: the hand-wave 'you get the idea' (or point, picture, gist), said to the reader to leave the rest of a list unsaid. Its value climbs word by word and peaks on the final noun. Near variants, translations and 'and so on' get a weaker…","st":"tested","i":6,"c":7,"m":[["core",12.1],["core: point",11.6],["future tense",9.1],["third person",6.1],["other verb",3.7],["done (neg)",3.7],["list tail",3.6],["control",1.5],["control",1.0]]},{"id":"L6_E56_N104","k":"m","L":6,"E":56,"N":104,"t":"A neuron for the slot naming what you're given","l":"This layer-6 neuron marks the word that names material being handed to the reader. In task instructions ('you are given a ___'), the word after 'a' fires almost regardless of what it is. In reading questions, the word after 'Based on this' fires, most strongly for 'passage'. Its…","st":"tested","i":5,"c":7,"m":[["presented-a short",0.7],["shown-a prompt",0.5],["given-a public",0.4],["the passage",0.2],["this passage",-0.0],["she-presented",0.0],["control: weather",0.0],["control: results",0.0],["Indications are",0.0],["Signs point",-0.5]]},{"id":"L7_E105_N158","k":"m","L":7,"E":105,"N":158,"t":"Data keys versus field labels: a signed key-delimiter neuron","l":"This neuron reads the delimiter after a key in a code literal and signals what kind of key it is. When the key is data, such as a variable, a name or a date inside a map being built, it fires positive. When the key is a record's field label, especially a labeled argument in a…","st":"tested","i":5,"c":6,"m":[["proper-noun key",3.0],["variable key",2.6],["date key",2.0],["field label",0.6],["tuple comma",0.5],["control math",0.3],["control call",0.0],["constructor noun",-1.9],["field quote",-2.0],["constructor label",-4.6]]},{"id":"L7_E13_N495","k":"m","L":7,"E":13,"N":495,"t":"A neuron that marks the moment symbols get declared","l":"This neuron responds when a problem lists single-letter symbols and then declares them, as in 'Let a, b, c be three positive real numbers.' Its value drops along the list and is most negative on the declaring words 'be', 'three' and 'numbers'. Logic and geometry letters work…","st":"tested","i":5,"c":7,"m":[["one symbol",0.0],["names",-0.0],["English noun",-0.0],["plain phrase",0.0],["four symbols",-0.3],["geometry points",-1.1],["LaTeX symbols",-1.2],["logic letters",-1.2],["a, b, c",-1.7]]},{"id":"L7_E154_N44","k":"m","L":7,"E":154,"N":44,"t":"A neuron that recognizes company names mid-spelling","l":"This neuron fires at the exact subword where a memorized company name, spelled across several tokens, becomes recognizable: the 'olia' of Veolia or the 'IS' of STERIS. It favours industrial, utility and consulting firms. It is quiet for people, places, single-word brands, and…","st":"tested","i":6,"c":8,"m":[["core B2B",10.2],["core caps",8.9],["word completion",6.0],["core agri",5.6],["same piece",0.1],["halo firm",-0.0],["one-word firm",0.0],["person",-0.0],["media brand",-0.5]]},{"id":"L7_E162_N120","k":"m","L":7,"E":162,"N":120,"t":"A Chinese neuron that tells story actors from opened topics","l":"This neuron in Qwen3.6 reads classical Chinese nominals ending in 者 and signs them by the role they play. When 'those who drink' are actors in an ongoing story, it fires strongly positive. When the same kind of noun opens a definition, a sage's saying, or a modern biography ('Li…","st":"tested","i":5,"c":6,"m":[["narrative 而",3.5],["narrative",3.3],["narrative 唯",3.0],["modern prose",2.2],["code control",2.0],["list control",1.0],["topic 夫",0.7],["dictum",-0.4],["dictum 故",-0.5]]},{"id":"L7_E169_N153","k":"m","L":7,"E":169,"N":153,"t":"The right one for you: a neuron for personal fit","l":"This neuron fires when a sentence about finding the thing that fits someone ends on 'you' or 'me': 'find the right course for you', 'el mejor patito para mí', 适合您的. The suitability frame alone gives only a weak signal. The pronoun for the speaker or listener is what opens the…","st":"tested","i":5,"c":8,"m":[["beginners",-0.0],["free (control)",-0.8],["her",-2.2],["him",-3.0],["yourself",-4.3],["everyone",-4.3],["them",-4.5],["us",-5.6],["me",-6.4],["you",-8.6]]},{"id":"L7_E240_N333","k":"m","L":7,"E":240,"N":333,"t":"A neuron for 'items to its left that…'","l":"This neuron fires on the word just after a phrase that places a set of items on one side of an anchor, as in 'two elements to its left that' or 'four people ahead of him,'. It is at home in array, queue and seating puzzles. Its value builds through the phrase and peaks at the…","st":"tested","i":5,"c":8,"m":[["to its left",8.1],["to its right",7.7],["after it",6.2],["before it",6.1],["near it",5.3],["larger than",5.0],["left (remaining)",3.2],["like it",2.0],["in the list",1.4]]},{"id":"L7_E252_N233","k":"m","L":7,"E":252,"N":233,"t":"A neuron that closes 'the lengths they will go'","l":"This expert neuron in layer 7 marks the word that closes a fronted effort idiom, the 'go' in 'the lengths they will go', reaching about +1.2. It also fires on a rating slash ('4 / 5') and on pieces of technical identifiers. Its negative side is the x in polynomial rings like…","st":"tested","i":4,"c":4,"m":[["lengths…go",1.5],["they will go",1.2],["extremes…go",-0.0],["how far…go",-0.0],["other verb",-0.0],["endure",-0.0],["places (control)",-0.0],["gym (control)",-0.0],["ring [x]",-0.4]]},{"id":"L7_E98_N259","k":"m","L":7,"E":98,"N":259,"t":"A letter-string lookup: Osch|man fires, hemangi|oper goes negative","l":"This neuron works like a lookup table of exact letter strings, not a concept detector. It fires most on 'man' after an o/e+ch piece (Osch|man, Och|man), even for invented names, while the real surnames Hochman and Kochman stay silent. A separate list goes negative: the 'oper' in…","st":"tested","i":3,"c":2,"m":[["Osch|man (core)",1.9],["Esch|man",1.0],["Lech|man",0.4],["Asch vowel (control)",0.2],["Och|man (invented)",-0.0],["place piece",0.0],["real name (control)",0.0],["other ending (control)",0.0],["tumor piece (neg)",-1.3]]},{"id":"L8_E119_N189","k":"m","L":8,"E":119,"N":189,"t":"A mining expert's place reader: Vietnam and orogens in, makers out","l":"Inside a layer-8 expert that the router reserves for mining and earth-resource text, this neuron marks where things are: tectonic belts like 'orogen', Vietnam and the other Asian and African gem countries, and region nouns like 'district'. It turns negative when ad copy credits…","st":"tested","i":4,"c":6,"m":[["Vietnam (core)",1.7],["region noun",1.5],["Myanmar",0.9],["Tanzania",0.7],["Thailand",0.5],["Russia",0.2],["control",0.2],["Western country",0.1],["control",0.1],["maker (neg)",-0.1]]},{"id":"L8_E134_N26","k":"m","L":8,"E":134,"N":26,"t":"A neuron that anticipates animals scent-marking their territory","l":"This neuron goes strongly negative when an animal is about to leave urine or scent to claim territory. It peaks on the verb 'mark' (or 'marcando', 'markieren') once the context makes marking likely, reaching about -14. Smell alone, territory alone, or rivalry alone give weaker…","st":"tested","i":7,"c":7,"m":[["songbirds (weak)",-7.6],["bees (halo)",-7.9],["teachers (control)",-8.9],["chemists (control)",-9.8],["deer",-10.1],["toddlers (control)",-11.0],["cats",-11.3],["dogs (core)",-11.5],["chimpanzees",-11.5],["wolves",-11.6]]},{"id":"L8_E145_N383","k":"m","L":8,"E":145,"N":383,"t":"First dash of a config list, and Dev Container 'features'","l":"This neuron marks the first dash that opens a new list in container and CI config files (docker-compose ports, GitHub Actions branches, Kubernetes ports). Its strongest values come from the word 'features' in VS Code Dev Container text. Within the same Docker topic, command-line…","st":"tested","i":3,"c":4,"m":[["depends_on: first dash",2.5],["networks: first dash",2.2],["ports: first dash",2.1],["later dash",2.1],["off-topic key",1.5],["environment: argv key",0.9],["control: comment",0.1],["control: scalar",-0.1],["version digit",-2.3],["docker ls",-4.8]]},{"id":"L8_E167_N367","k":"m","L":8,"E":167,"N":367,"t":"A neuron for driven, steered wheels and counted axles","l":"This neuron in an early-middle layer fires negative when a vehicle part is given a mechanical role: wheels that are 'driven' or 'steered', parts that are 'steerable', and axles that are counted or multiplied ('three axles', 'multi-axled'). The router does most of the sense…","st":"tested","i":5,"c":6,"m":[["control: color",-0.1],["offset (halo)",-0.4],["control: new",-0.5],["single axle",-3.0],["counted axles",-3.3],["driven",-3.9],["tandem axles",-5.3],["steerable",-5.5],["undriven",-6.6]]},{"id":"L8_E235_N17","k":"m","L":8,"E":235,"N":17,"t":"A neuron that anticipates 'trachomatis' after Chlamydia tr|ach","l":"This neuron predicts the bacterium Chlamydia trachomatis. When the pieces ' tr' and 'ach' come right after the word Chlamydia, it fires on 'ach' about ten times higher than almost anything else, before the rest of the name appears. Each cue alone gives a weak signal: the genus…","st":"tested","i":4,"c":10,"m":[["Chlamydia",10.2],["C. abbrev",9.4],["all caps",9.3],["misspelled",8.7],["Neisseria",6.4],["other genus",6.1],["Streptococcus",3.6],["no genus",0.9],["rock control",0.8]]},{"id":"L8_E23_N349","k":"m","L":8,"E":23,"N":349,"t":"A take_action and subparsers neuron, negative on OpenStack infrastructure","l":"This neuron responds to the vocabulary of Python command-dispatch code in coding tasks: the name take_action, Command classes, subparsers and talk of extending a CLI. It looks like memorized co-occurrence with frameworks such as cliff, argparse and Django, not a general idea of…","st":"tested","i":4,"c":7,"m":[["take_action",4.2],["take_command",1.3],["take_move",1.3],["perform_action",0.8],["get_action",0.5],["list_commands",0.5],["on_action",0.3],["control",0.1],["Neutron NAT",-1.7]]},{"id":"L8_E29_N120","k":"m","L":8,"E":29,"N":120,"t":"Software feature nouns up, Evernote's 'note' sharply down","l":"This neuron responds weakly to the generic nouns software is built from, such as backup, filters, routing, project menu and Metadata, most of all inside feature lists. Its deep negative response is a lexical reflex: the token 'note' completing 'Evernote' drives it to about -12…","st":"tested","i":3,"c":7,"m":[["UI noun",2.2],["feature item",2.1],["tool noun",1.5],["metadata",1.5],["halo",0.2],["control",0.0],["control",-0.1],["other brand",-0.4],["OneNote",-5.2],["Evernote",-9.6]]},{"id":"L8_E31_N374","k":"m","L":8,"E":31,"N":374,"t":"A neuron for quoted classical verse: Latin poetry and Sanskrit","l":"This neuron responds to authentic classical-language writing quoted inside modern text, chiefly Latin poetry and Sanskrit verse, with peaks several words into the passage. For Sanskrit, the router does the filtering: it admits Sanskrit but rejects modern Hindi written in the…","st":"tested","i":7,"c":8,"m":[["Ovid verse",3.0],["Sanskrit (Devanagari)",2.3],["modern Hindi",2.2],["romanized Sanskrit",2.0],["Virgil verse",1.4],["Indic names",0.3],["Latin epitaph",0.2],["math",0.2],["Caesar prose",0.0],["plain English",0.0]]},{"id":"L8_E55_N507","k":"m","L":8,"E":55,"N":507,"t":"A neuron that hits its floor on 'four-stroke'","l":"This neuron tracks the piston engine's four-phase cycle, and its meaningful direction is negative. Narrating intake, compression, power and exhaust drives it down, and it bottoms out near -16 on the word that names the cycle: 'four-stroke', 'cuatro tiempos', 'vier Takte'. Heat…","st":"tested","i":6,"c":10,"m":[["heat pump",-0.7],["gas noun",-1.6],["control",-2.2],["Carnot",-4.4],["control",-5.4],["phase word",-6.6],["Otto",-6.9],["two-stroke",-7.5],["six-stroke",-11.0],["four-stroke",-13.2]]},{"id":"L8_E58_N93","k":"m","L":8,"E":58,"N":93,"t":"A contract neuron that marks who or what caused it","l":"This neuron goes negative on the word that fills a cause slot in a contract: 'if caused by…', 如因…原因, 'causes beyond its control'. It goes deepest when the cause belongs to one of the contracting parties, reaching about -7 on 如因乙方自身原因. Whether the outcome is good or bad makes no…","st":"tested","i":6,"c":7,"m":[["control: duty",0.2],["control: address",0.2],["control: copies",0.1],["weather cause",-3.5],["bare 'if due to'",-3.5],["bare 'because'",-3.5],["beyond control",-4.4],["party cause",-5.8],["other party",-6.0],["party's own cause",-7.3]]},{"id":"L8_E65_N282","k":"m","L":8,"E":65,"N":282,"t":"A neuron for the moment effort 'went into' something","l":"This neuron fires strongly negative on the verb that says effort went into something: 'went', 'has gone', 'put', or German 'gesteckt', when the thing moving is effort, thought, work or care. It fires before 'into' appears. Other abstract or physical things that 'went' (luck…","st":"tested","i":5,"c":8,"m":[["rain (control)",-4.3],["hatred (abstract)",-5.0],["money (floor)",-6.7],["love",-7.2],["thought",-8.6],["time",-9.3],["care",-9.8],["research",-10.3],["effort",-11.9],["work",-12.1]]},{"id":"L9_E172_N164","k":"m","L":9,"E":172,"N":164,"t":"A Unix neuron that dives on block and character device files","l":"In Unix text, this neuron tracks what kind of file a token names. It fires strongly negative when the file is a block or character device, whether marked by the b or c letter in an ls -l listing, the words 'block' and 'character', or the -b/-c shell tests. Sockets, pipes and…","st":"tested","i":6,"c":8,"m":[["non-empty file",0.9],["regular file",0.0],["empty string",-0.0],["directory",-0.1],["named pipe",-0.3],["readable",-0.4],["socket",-1.3],["char device",-3.6],["block device",-4.5]]},{"id":"L9_E183_N305","k":"m","L":9,"E":183,"N":305,"t":"A neuron for hibernate, swap and distro names in Linux installs","l":"This neuron fires on the names of low-level system objects when someone is installing or running a PC operating system. The strongest case is 'hibernate' or 'swap' while partitioning a Linux disk. Distro names, live and rescue media, and GPU vendors in driver instructions also…","st":"tested","i":4,"c":6,"m":[["hibernate",2.9],["power state",1.6],["boot media",1.5],["GPU driver",1.3],["swap",1.2],["distro",1.2],["halo",0.1],["collective noun",-0.0],["control",-0.0],["control",0.0]]},{"id":"L9_E18_N82","k":"m","L":9,"E":18,"N":82,"t":"A JSON-LD detector with a faint HTML <head> echo","l":"This neuron is a JSON-LD detector. It goes deeply negative on the '-L' after 'JSON', and SEO or product-page context pushes it lower still. Weaker markup words such as schema markup, Microdata and script tags pull it partway down. Its positive side is roughly ten times weaker…","st":"tested","i":3,"c":8,"m":[["<head",0.4],["photo control",-0.1],["price control",-0.1],["schema markup",-0.6],["script tag",-0.8],["structured data",-1.2],["JSONL control",-3.8],["YAML-LD",-4.4],["JSON-LD (article)",-6.8],["JSON-LD",-7.0]]},{"id":"L9_E254_N135","k":"m","L":9,"E":254,"N":135,"t":"A neuron for variables that belong to load-testing tools","l":"This neuron responds to the word 'variable' when the text is about a specific testing or authoring tool that keeps its own variables, above all JMeter, LoadRunner and Postman. It also fires on the calls that save values into such stores. The same word in an ordinary coding…","st":"tested","i":5,"c":6,"m":[["JMeter",11.1],["Gatling",9.2],["Postman",7.9],["LoadRunner",6.7],["Storyline",5.1],["Jenkins (control)",4.1],["Excel (control)",3.7],["Python exercise",3.0],["coding problem",2.8],["Ansible",2.6]]},{"id":"L9_E32_N122","k":"m","L":9,"E":32,"N":122,"t":"A neuron that gives a fixed sign to a few recognized words","l":"This neuron gives a fixed sign to a handful of multi-token words once the model has recognized them. It is strongly positive on the identifying piece of Smoot, Nottoway, tomahawk and the Spanish cake-tester word palillo, and negative on Stroganoff, Hadley and Qt 'Ui_' class…","st":"tested","i":2,"c":1,"m":[["Reed Smoot",0.7],["Smoot (core)",0.6],["Hiram (halo)",0.3],["Borah (halo)",0.2],["Norris (control)",0.1],["Wheeler (control)",0.1],["Snoot (letters)",0.0],["Hadley (neg)",-0.0],["Stroganoff (neg)",-0.1]]},{"id":"L9_E75_N365","k":"m","L":9,"E":75,"N":365,"t":"A neuron for the Family and Medical Leave Act and unpaid leave","l":"This neuron recognizes the US Family and Medical Leave Act. It fires hardest on the law's name, including misspelled or translated versions. Without the name, it still responds to the wording of a leave grant: employees may take a number of weeks of unpaid, job-protected leave…","st":"tested","i":5,"c":8,"m":[["acronym",12.1],["unpaid leave",7.5],["statute name",7.1],["job-protected",7.0],["military leave",1.7],["sabbatical",0.7],["sick leave",0.5],["control",-0.1],["decoy control",-0.2],["paid vacation",-0.5]]},{"id":"L9_E94_N72","k":"m","L":9,"E":94,"N":72,"t":"A commitment dial at clause ends: half-granted versus fully vouched","l":"This neuron reads how fully a writer commits to a claim, at the punctuation that closes it. A half-hearted grant such as 'might be fine for a dim room,' 'is okay,' or 'well, almost –' pushes it positive. Full commitment such as 'for sure,' 'certainly,' 'I agree that…' or…","st":"tested","i":7,"c":8,"m":[["modal hedge",2.6],["lukewarm",2.1],["seems",0.9],["plain praise",0.7],["for sure",0.6],["negated modal",0.5],["unmarked",0.1],["factual control",0.1],["certainly",-2.3]]}]