{"path":"research/univocity-and-the-concept-layer.md","content":"# Univocity and the concept layer: one sense in the record, many senses in the graph\n\n**Date**: 2026-11-06. **Type**: a session's reading, for his review of the implications; nothing here is\nruled except the rows and lines it cites as ruled. **Ledger**: BH7 (open). **Home of the global rule it\nstarts from**: `~/.claude/CLAUDE.md`, *Univocity in what Fredrik reads*; its evidence,\n`~/.claude/directive-evidence.md` § The established term.\n\n## 0. In short\n\nThe rule made for chat this week has two names, one for each pole. **Univocity** is terminology\nscience's ideal that one term designates one concept (ISO 704 § 7.7, *monosemy*, and its twin\n*mononymy*, one name per concept; Wüster's general theory of terminology). **Plainwashing**, his\ncoinage of 2026-11-05, is the writer's swap of an established term for a plainer, vaguer everyday word\nthat reads as kinder and says less (*law* for *strict generalization*; *the room* for the AI Safety\nDay demo; *the hits that cost something* for *discriminating cases*). Plainwashing is the writer's\nroad into **equivocation**, the fallacy in which one word carries two senses across a stretch of\nreasoning, and equivocation is the shape of every long-lived mistake on this project's record\n(`the-kind-of-mistake-we-make.md` § 3: parts taken for reasons, polarity for direction, a premise for a part).\n\nHis ask, 2026-11-06: *\"the whole univocity/plainwashing thing relates deeply to Deliberus AND to the\nconcept layer inside Deliberus. Document this carefully for us to review implications later.\"*\n\nThe reading, in one sentence: **the project's own words must be univocal, and the graph's contents\nmust not be**, since the concept layer exists to hold the many senses of a contested word as\nattributed uses; the two live at different levels, and each of the two ways of collapsing them is a\nmistake the corpus has already met.\n\n## 1. His words\n\n- 2026-11-06: *\"it HELPS me to understand what is meant when a SPECIFIC word with a SPECIFIC meaning is\n  in use. Then my brain just runs like a train on tracks that I can see before me, if ambiguity is\n  introduced it's like I'm in thick fog and am experiencing my own little version of the trolley\n  problem, dunno what mental tracks to toot forward to.\"*\n- 2026-11-06, on the swap in a research note: *\"Can you see how stuff like 'the three that cost\n  something' is vague and cryptic and sorta plainwashing to someone (like me) who hasn't been there to\n  see all the work fly by of researching/reading/analyzing/coding and who just sees the summary at the\n  end of the turn\"*\n- 2026-11-05: *\"I'm very annoyed at muddling ambiguous words like 'law' being used if there's already an\n  established technical term that we've begun using.\"*\n- 2026-10-22: *\"Why use the word line and rope?!?! You use so many different words to refer to the same\n  thing. Stop being so goddamn poetic all the time, this is what's causing the muddling I've been\n  sensing and complaining about for some time now.\"*\n- 2026-10-18: *\"I have a hunch using words a bit too interchangeably and haphazardly is what caused our\n  confusion around parts/reasons months ago. I went along with you using 'parts' as stand-in for reasons\n  and didn't realize the meaning had shifted/drifted.\"*\n- 2026-10-18, the other half: *\"any claim that uses a concept also simultaneously argues about what it\n  means? I mean, that's also how natural language behaves, so I guess it works?\"* (ledger AM7, a use is\n  a vote on the meaning, ruled 2026-10-19).\n- 2026-08-26: *\"I'm not against jargon when I know what it means.\"*\n\n## 2. What the project already holds on the record's own words\n\nEach of these is univocity under another name, applied to the record (notes, prompts, the ledger,\nthis file's language, chat with him):\n\n| Where | What it says | Univocity's part |\n|---|---|---|\n| The analytic register (`CLAUDE.md` § Critical Design Philosophy, blessed 2026-10-15) | a term is defined before it is used, one word does one job across the corpus | monosemy as a working rule |\n| BD11, a concept takes its discipline's name (ruled 2026-10-25) | about to coin a project word for a concept a discipline studies, propose the discipline's term | mononymy, and the term's source |\n| BD12, layers for a spectrum, two names for two things (ruled 2026-10-25) | a colliding word is decomposed into layers, or renamed when its senses are two things | the repair when univocity has already failed |\n| BD13, two things are called one only when every operation applies to both (ruled 2026-10-25) | *is just*, *the same as*, *one structure* are tested against what the project does to each | the test for a false merge of concepts |\n| The four-way naming rule (`CLAUDE.md` § READ FIRST, 2026-10-19) | *link*, *reading*, *check*, *proposal*, *version*, *review* each say which of their jobs they do in the sentence | univocity per sentence where mononymy is lost |\n| The glossary's retired words (*step*, *the number*, *line*, *rope*, *machine pass*) | replaced by the exact term, never glossed | a retired synonym is mononymy enforced |\n| The three headline tests (ruled 2026-08-29; the wanted-word rider 2026-10-07) | stranger: does the word say what it means to someone without the context; collision: does it already do another job here | the collision test is a univocity check at the moment of coining |\n| AM12, a pass is taught the method, never the graph's contents (ruled 2026-10-24) | a microtask's grounding carries definitions | univocity in the prompts |\n\nSo the record's side is ruled, piece by piece. What the global rule of 2026-11-05 adds is the\n*writer's* cue (the word about to leave the text, in chat and in notes) and the name of the pull the\nother way. The cost of failing it is measured: every one of the six costliest mistakes in § 3 of\n`the-kind-of-mistake-we-make.md` was a word carrying two things, and none threw.\n\n## 3. Where the concept layer stands, and why it is the other pole on purpose\n\nThe concept layer is built on the fact univocity denies for the record: a contested word has no one\nsense. Its rulings and findings, all already on the record:\n\n- **A word's core is shared, its periphery is assembled per community and per person, and for a\n  contested word the periphery is the meaning** (`core-periphery-and-the-concept-layer.md` § 5; Marti\n  et al. 2023: two random people match on a word's whole profile 12 to 33 % of the time, and assume two\n  thirds agree).\n- **The layer records deltas, never descriptions** (§ 5b, his sharpening of 2026-09-06): a sense is\n  somebody's difference from a shown reference, and the centroid (the model's reading of the population's mean)\n  is demoted from a sense anyone holds to the origin deltas are measured from.\n- **A use is a vote on the meaning** (AM7): every claim using a concept is one attributed use; a\n  definition is a claim on the word; the layer is read over uses and never written to directly.\n- **Nobody owns a wanted word** (AC1, AC2, ruled 2026-10-07): how much a word is wanted (*rational*,\n  *truth*, *AI*) is a property beside its senses, a move on it is a claim of a named kind, and the\n  same content under a different standing is a sixth way senses differ.\n- **The motte and the bailey** (`the-motte-and-the-periphery.md`): the retreat from a persuasive\n  periphery to a defensible core under challenge is a live re-assembly of the concept, which the layer\n  would misread as polysemy until sense attribution is stored; four form tests tell refinement from the\n  manoeuvre.\n\nSo the graph's words are polysemous **as data**: the layer's whole job is to keep the senses apart and\nattributed, which is the opposite of forcing one. A single sense per word in the graph would be the\nsixth flattening mechanism named in that note, **centroid sense minting**: the average becomes the\nbearer of the word's meaning, nobody's sense stands, and the disagreement the graph exists to map is\nerased before it is measured.\n\n## 4. The two levels, and the two collapses\n\n| Level | What univocity demands | What happens if the levels are confused |\n|---|---|---|\n| **The record about the graph**: notes, ledger, canon, prompts, chat with him, the glossary | one term, one sense, the discipline's; synonyms retired; a coinage glossed or replaced | letting the record's words carry two senses is the class of parts taken for reasons: silent, long-lived, expensive (§ 3 of the mistakes note) |\n| **The graph's contents**: claims, concepts, senses, uses | nothing; the layer stores every attributed sense and reads the spread | forcing one sense is centroid sense minting, flattening mechanism six; the wager's instrument goes blind on the words where disagreement lives |\n\nThe ontology note's sorting test says the same thing from its own side (`what-belongs-in-the-ontology.md` § 3: *if two users disagreed about this, could the graph represent both of their positions?* Yes, it is a claim; no, it is ontology, load-bearing, to be made visible). Two users cannot disagree inside the graph about what *link* means in the keeper's prompt or what *reason* means in a ledger row: those words are ontology, so they are held to one sense and made visible (the glossary, the four-way naming rule). Two users can disagree inside the graph about what *justice* or *intelligence* means: those words are claims, so the graph holds every sense and lets them be attacked. Univocity is the ontology side of that test; the concept layer is its claim side.\n\nThe project already lives on both levels at once without naming it: the four-way naming rule holds\n*link* to one job per sentence in our prose, while the graph stores the same word *link* as whatever a\nsource's author meant by it, attributed. What this note adds is the name for why both are right, and\nthe two places where they touch.\n\n## 5. Where the two levels touch, and the implications to review\n\n### 5a. Extraction and the splitter can plainwash an author, and most of the guard is built\n\nA claim is cut for the stranger: decontextualised, in plain words, so that a reader without the\nsource follows it. That is the right rule for a claim's *core*. But the author's own term is a\n*use*, one attributed vote on the word's meaning (AM7), and a rewrite that replaces *strict\ngeneralization* with *rule* deletes the vote and files the author under the centroid. Read against\n`a-sentence-as-a-pointer-into-a-landscape.md` § 6 (whole, this session), the guard is further along than\nthe chat rule: AA6's instruction half, ruled and built 2026-10-19 (*\"Yes!\"*), tells pass 2b to change\nonly what a stranger needs, keep the author's words for the rest, put added context in square\nbrackets, and add no weight or certainty; each claim keeps the author's assertion as `author_wording`;\ncode counts the content words a rewrite added (`words_added`, `words_added_in_brackets`); and § 6d\nreads a bracketed gloss that defines a term as the author's unexpressed definitional commitment,\nattributed to the author and marked reconstructed, *a use, never the meaning* (the storage law, AA1:\nread, never set). So plainwashing an author is already named there, under other words.\n\nWhat is not yet caught is the exact case of this note: a **replaced** term. The mark sees a word\nfound nowhere in the source (*persistent* for *second*), not a word that stood in for the author's\nterm and happens to occur elsewhere in the source. **Implication 1, to review**: the wording pass\ngains a count of terms replaced, checkable by code against the passage (a term in the passage absent\nfrom the claim, with a plainer word in its position), and a claim that drops a term the source uses as\na term owes a reading. The decision-model check on added words (§ 6c there, built and tried) is the\nshape it would take.\n\n### 5b. The stranger test and univocity are assigned to audiences, not opposed\n\nP28 (words for everyone who might come, ruled 2026-10-23) governs what a visitor reads; the\nconstellation's stranger test governs a tile; univocity governs what he reads and what the record\nsays. These do not conflict: the audience sets which pole is kept, and the thing that must never\nhappen is **the same word doing both jobs in one text**, a term for him and a plain word for the\nvisitor, unmarked. **Implication 2, to review**: the headline tests gain the audience as an explicit\nfirst question (for whom is this line), since the stranger test applied to a resident line or a\nledger row pulls it toward plainwashing, and univocity applied to a landing tile pulls it toward a\nword that must be learned (condition (d) of the success prior).\n\n### 5c. The glossary and the concept layer are the same object at two scales\n\nThe blessed glossary (`~/.claude/glossary/<project>.md`) is the record's register of univocity: one\nsentence per term, one status per reader (him). The concept layer is the graph's register of\npolysemy: senses per community, attributed, with lifecycle states. Both store *a word, what it means\nhere, and for whom*. The glossary has one reader and one sense per term; the layer has many of each.\n**Implication 3, to review, and only a thought**: a reader's *known / not-yet* on a term is a\nreader-relative fact of the same kind the layer stores for a community's sense; if Deliberus ever\nshows a visitor the project's own terms (the ledger page, the docs), the glossary is the first row of\na concept layer with one row per reader, and the wheel built today (`glossary-ui`) is its first reading surface.\nThis is not a build proposal; it is where the two meet. *Built the next day, 2026-11-07, on his word\n(\"build this in/weave it into my vocabulary/glossary/glossing practice ... as a globally available and\nproactive tool\"): `~/.claude/glossary/global.md` filled and read beneath every project glossary, and a\n**stand-ins table** at the end of each glossary listing the plainer words that have stood in for a term,\nwith where each was seen; the counter and the docs nudge flag a stand-in used where the term is absent.\nThe stand-ins table is the glossary's own delta register in miniature: per term, the swaps one reader has\nmet, which is the shape § 5c describes for the concept layer with many readers.*\n\n### 5d. The microtasks: a prompt's words are the pass's error rate\n\n*An error rate belongs to the task as written before it belongs to the model* (ruled 2026-09-29):\nof nineteen measured failures of microtasks, seventeen traced to how the question was posed\n(`pass-quality-and-the-choice-of-model.md` § 3). A plainwashed prompt is the cheapest way to pose a\nquestion wrongly: *reason* for *premise*, *part* for *reason*, *the number* for *posterior* in a\nprompt gives the model two tracks exactly as it gives him two. **Implication 4, to review**: the\ngrounding block of every pass (`deliberus/grounding.py`) is read for retired words and unglossed\ncoinages the way `glossary-docs-nudge` reads a note, by code, at render time; a pass whose prompt\ncarries a retired word is outdated the way a changed instruction outdates it.\n\n### 5e. The muddle class is the equivocation fallacy, and the graph can hold it as content\n\nThe record's worst mistakes were equivocations; the graph's method for a contested word is to store\neach sense and let the clash be argued (AM7: the verbal-classification scheme's critical questions\n*are* the meaning dispute). **Implication 5, to review**: when a session finds an equivocation in the\nrecord (a word that has done two jobs), the repair is BD12 (layers or two names) for the record, and\nthe same pair of senses is a candidate for the concept layer as a worked example of what it holds,\nsince the project's own history is the case documented best of the mechanism it instruments. This\ntouches the question of operator reflexivity (the project hosting its own design in its graph; self-hosting, held by his word of 2026-08-30), so it waits\nwith that hold.\n\n## 6. What this does not claim\n\n- It does not claim that technical terms are always better than plain words for every reader: P28\n  and the stranger test stand for visitors; univocity is for him and for the record.\n- It does not claim the concept layer should hold the project's own vocabulary: § 5c is a thought\n  for review, and self-hosting is held.\n- It does not measure anything new: the counts cited (nineteen failures, six costliest mistakes,\n  Marti's 12 to 33 %) are the notes' own, read whole this session where cited by section.\n\n## 7. How this was done\n\nRead whole this session: his messages of 2026-08-26, 2026-11-05 and 2026-11-06 from the session\ntranscripts; `a-sentence-as-a-pointer-into-a-landscape.md` § 6 (6a to 6e); `what-belongs-in-the-ontology.md` § 3; `store-the-uses-read-the-meaning.md` § 0 and its row AA1 (the storage law); the ledger rows BD11, BD12, AM7, AC1, AC2, BH3, BH4; `the-kind-of-mistake-we-make.md`\n§ 3; `core-periphery-and-the-concept-layer.md` § 5b; the univocity rule and its evidence section; the\nPlain Brief words rule; the glossary calibration doc of 2026-08-26 (§ 2 and the paragraph on technical\nvocabulary). Not read whole: the three notes above beyond the sections named, `jev-and-the-judgement-layer.md`, `prime-real-estate-words.md`, `the-motte-and-the-periphery.md`,\n`pass-quality-and-the-choice-of-model.md` beyond its § 3 figure; their claims above are taken from the\nTopic Index lines and the ledger rows that cite them. No outside research beyond one search confirming\nISO 704's monosemy section. Helpers: none.\n\n## Cross-references\n\n- `core-periphery-and-the-concept-layer.md` § 5, § 5b, § 6.5 (the sixth mechanism)\n- `the-kind-of-mistake-we-make.md` § 3, § 10a\n- `the-motte-and-the-periphery.md` § 3b, § 3c\n- `prime-real-estate-words.md` § 4, § 5e\n- `pass-quality-and-the-choice-of-model.md` § 3\n- `docs/ux-principles.md` P28; `specs/landing-redesign/constellation-map.md` (the stranger test)\n- `~/dotfiles/docs/glossary_and_jargon_calibration_2026-08-26.md`\n- ledger: AM7, AC1, AC2, BD11, BD12, BD13, AA1, AA6, AM12; BH7 (this note's row)\n- `a-sentence-as-a-pointer-into-a-landscape.md` § 6c, § 6d (the instruction half of AA6; a gloss as the author's commitment)\n- `what-belongs-in-the-ontology.md` § 3 (the sorting test)\n"}