{"path":"research/exact-terms-audit.md","content":"# The exact-terms audit: each microtask's prompt, fields and code read for stand-ins\n\n*Opened 2026-11-07. His question that day: \"Would you say the exact-terms audit could be good to have done?\" and\nhis word after the answer: \"let's start with the audit via the splitter\". The ruling it applies: a concept takes the\nname its discipline already gives it (ledger BD11), one word does one job (`the-kind-of-mistake-we-make.md` § 3),\nand the univocity directive (a term, glossed once; never a plainer word in its place). The order it keeps: one\nmicrotask at a time, each audited as the first step of its review by a person (Y9, a pass is read by a person before\nit is proven), before his cases of it are read, since a wording change outdates every earlier reading of that\nmicrotask (T8) and no microtask holds a proof of its current version, which is why the audit costs least now.*\n\n## 0. Method, and what a finding is\n\nRead whole: the prompt as sent, the system turn, every field name and description in the answer models, the\nenum values, and the module's own words where they name the same things. For each word that stands for\nsomething the discipline or the glossary already names, one row: the word, the sentence it sits in, the term,\nand whether the swap changes only the words or the question. A row that changes the question is put to him as\nits own decision; the rest are proposed as a batch for one blessing. A collision (one word doing two jobs across\nthe code) is a row too, with the other job named, and it is settled by BD12 (a colliding word is decomposed\ninto named layers when its senses form one scale, and renamed when they are two different things). Each\nmicrotask's answers keep one definition, identical in its instructions, on any review page and in the code that\nacts on them (T5, ruled for the keeper's seven keys; the field names below are the splitter's keys).\n\nThe home of each microtask's audit is a section here; the glossary's stand-ins table gets a row for a stand-in\nthat has also reached chat or a note.\n\n## 1. The splitter: `split.propose` and its check `coverage.check` (`deliberus/extraction/decompose_claim.py`)\n\nRead 2026-11-07, whole: prompt `_PROMPT` with `_THEORY`, `_COVERAGE_PROMPT`, the two system turns, the\nmodels `DecomposedPart`, `DecompositionProposal`, `PartCheck`, `CoverageReport`, the `AXES` tuple, and the\nreaders of the fields (`deliberus/proposed_split.py`, `scripts/recut_splits.py`, the web app's split view).\n\n### 1a. The one row that changes the question (his decision)\n\n| word | where it sits | the term | what changes |\n|---|---|---|---|\n| **weaker** | the part test, `if_false: weaker`, and the prompt's \"if this part were false, would the claim be weaker?\" | **false** (*if this part were false, would the whole be false?*), with **lost support** as the second answer | The glossary's own row for *part* reads: *if it fell, the claim would be false, not merely less supported* (`a-part-and-a-reason.md` § 0), and § 11.2 of that note proposes exactly this split of the gate's word: *the claim would be false* makes a part; *the claim would have lost support* makes a reason. As worded the gate admits a reason (a claim whose reason fails is weaker too). Replayed 2026-10-12 it still sorted 39 parts right, so the fault is in the word, not yet in a measured outcome; but it is the word the reader is told to apply. The same split was proposed in `microtasks-in-broad-daylight.md` § 3h on 2026-10-12, and his word then was to wait: *\"Yeah sounds good but let's await the deep dive and what it gives back before we make any such change.\"* The deep dive is `a-part-and-a-reason.md`, and § 11.2 there gives it back as this row. Unruled refinement of T9. |\n\nA yes gives `IfFalseLit` the values `false`, `lost_support`, `unaffected`, `stronger`, `cannot_tell`, with\nonly `false` gating a part and `lost_support` filed as a reason for the whole (T9's second filing, which the\nsplitter cannot do yet). A no keeps *weaker* and the row closes.\n\n*Ruled and built 2026-11-11, his words after a catch-up: \"Yes, split it\". Splitter version 3:\n`false` and `lost_support` as above; `lost_support` holds the cut with its reason named, since the\nsplitter cannot yet file a reason; a part that repeats its whole holds it; the split sweep waits for\nreasons off on sentences that give reasons. What the live graph showed, and why he doubted the\nsplitter: `a-part-and-a-reason.md` § 11a. Bearing on this section since 7 November: BJ1 (one door\nsorts before it files; the claim page's box files a person's reason as a part, the same filing at a\nperson's door), and BL4 (the keeper's prompt changed, so its version moved and its held-out pairs\nneed a re-run before his reading).*\n\n### 1b. Words that change only the wording (proposed as one batch)\n\n| word | where it sits | the term | why |\n|---|---|---|---|\n| parent · original · the claim · the parent claim | throughout both prompts and all four models (four words for one thing) | **the whole** | AL15, the whole as a record; the glossary names *whole* and never *parent*. Wanner et al. say *parent*; our term is the whole, and one word. |\n| decompose · decomposition · cut · split | module name, prompt opening (\"Decompose this claim\"), the system turn, the glossary's own rows (*splitter*: \"cuts a bundled claim\"; AX8: \"split to the floor\") | **split** for the act; **decomposition** only where the discipline's measure is meant (coverage, coherence, atomicity, the axes) | Three words for one act. *Split* is the code's name (`split.propose`) and his own (\"Split split split to the floor\"). His call whether *cut* stays as his everyday word in chat; in prompts and notes it goes. |\n| seams (\"what kind of compound is this claim, and where are its seams?\") | `DecompositionProposal.reasoning` | **where a reader could affirm one part while denying another** (independent deniability, Russell's test, already in the prompt) | A metaphor standing in for the test the prompt states two paragraphs later. |\n| smuggled · `smuggled_value` · \"a smuggled evaluative modifier\" | field, `_THEORY`, the docstring | **a value premise carried by a modifier** (field: `fused_value_premise`) | *Value premise* is the project's term (`claim_type: value_premise`); *evaluative modifier* is Davidson's. *Smuggled* says the same with an accusation in it. |\n| POLARITY (\"KEEP EACH CLAUSE'S POLARITY\") | the prompt's rule heading | **what each clause asserts** (assertoric force, the discipline's word for whether a sentence puts its content forward as true; Potts's *at-issue* content, the content a hearer can answer *that's false* to) | Collision: *polarity* means support against attack everywhere else in the code (`relevance_judge.py`, `latent_questions.py`, the glossary's *polarity blindness*). The rule's own content is `held_as`: asserts, denies, concedes, supposes, absent, which is force, not polarity. The glossary's *polarity flip* (named 2026-09-29) is the same collision one level up; proposed rename *force flip*. |\n| grants · granted that · concedes | the prompt (\"only grants\", \"granted that Y\") against the enum value `concedes` | **concedes** (concessive clause) | Two words for one enum value; the enum's word wins. |\n| background it takes for granted | the prompt's polarity rule and the coverage prompt | **presupposition** | The axis is already named `presupposition` in `AXES`; the prompt describes it without naming it. |\n| scope (\"quantity or attitude separated from content\") | `AXES` | **quantifier scope** for the quantity; the attitude folds into **attribution** (a propositional attitude is who holds what) | Collision: *scope* is AT11's term for whether a claim speaks of one case or many (`scope_reading.py`), a different thing. |\n| already_atomic · \"the claim asserts ONE thing\" | field and prompt | **at the separability floor** (field: `at_the_floor`) | The canon line reads *nothing is permanently atomic*; a field that says a claim is already atomic asserts its opposite word for word. The prompt's test (asserts one thing) stays; the name changes. *Atomic* in the DnDScore sense (`atomic` against `decontextualized`) is the discipline's word for the part's minimal form and stays. |\n| value judgement (\"`unnecessary` is a VALUE judgement\") | `_THEORY` | **value premise** | The project's term, used three lines away. |\n| dropped · invented | `CoverageReport` fields and the coverage prompt's items 3 and 4 | keep the field names; the prompt names the axes once: dropped is a **coverage** failure, invented a **coherence** failure (Wanner et al.) | The two JHU axes are in the glossary and in the module docstring, never in the prompt. |\n| claim_type (empirical, normative, definitional, value_premise) | field and prompt | keep **type**; note the collision in the glossary | The glossary's row *claim kinds* lists these same values as the kind field, while BD6 (origin and role, told apart) splits `claim_kind` into origin and role, neither of which is this. Three fields, and the glossary conflates two. Not the splitter's fault; a glossary repair. |\n\nKept as the discipline's words, no row: proposition, compound, conjunct, truth-function (Russell's word for a compound whose truth is fixed by its parts' truth), modifier, event\nsemantics, rhetorical parallelism, causal chain, decontextualized, coverage, reconstructed (the recomposition\ncheck), supposes (AM5's *supposed* claim).\n\n### 1c. What the audit did not reach\n\nThe microtask's own quality (what it cuts well and badly) is the review proper, his cases one at a time; this\nsection only fixes the words it is asked in. The wording change, once blessed, is one version change\n(`docs/research/microtask-versions.json`), and every split read before it is marked outdated, which is what makes the\naudit cheapest before his reading, not after.\n"}