{"path":"research/corpus-graph/readers/2026-10-10-strength-semantics-and-schemes-as-a-reader-returned-it.md","content":"# What a reader found to bear on eleven notes on strength semantics, schemes and extraction research (2026-10-10)\n\nA read-only reader (Sonnet), one of those sent through the notes no reader had yet read, after the founder\nasked for the graph's coverage to be widened (README § 12). It was given the brief in\n[../reader_brief.md](../reader_brief.md) and these 11 notes:\n\n- `docs/research/qbaf-gradual-semantics-research.md`\n- `docs/research/the-irit-harvest.md`\n- `docs/research/listening-for-the-calculus-and-sturdy-microtasks.md`\n- `docs/research/perceiving-possibility-and-the-analogy-organ.md`\n- `docs/research/support-semantics-and-the-conjunction-problem.md`\n- `docs/inductive-reasoning-research.md`\n- `docs/research/walton-argument-schemes.md`\n- `docs/research/scientific-method-two-steps-and-deliberus.md`\n- `docs/research/extraction-improvement-research.md`\n- `docs/research/unbuilt-features-research-mar31.md`\n- `docs/research/definitions-as-alignment-tests.md`\n\nIt made 107 tool calls. Its report is below, unchanged. What was kept from it, after the quoted words were\nchecked by code and both passages of each finding were opened, is in `../found.jsonl` (lines whose `where`\nnames this file).\n\n---\n\n11 of 11 notes read in full.\n\n### F1\ntarget: docs/research/unbuilt-features-research-mar31.md:110\ntarget_words: \"Never auto-merge. The distinction between \"same conclusion, different evidence\" and \"true duplicate\" requires human judgment\"\nolder: docs/research/claim-sameness-philosophical-readings.md:7\nolder_words: \"anyone types anything, careful auto-merge runs afterward, and anyone can split\"\nkind: later-correction\nwhy: Founder ruling of 26 August 2026. Nothing-merges-without-explicit-action was rejected as friction at the wrong moment. Originals are never destroyed, merges are visible, splitting is a recorded act, the author is notified, the bar is higher for value and definitional claims, and an evidence gate precedes any switch-on, so nothing auto-merges yet. Someone building the dedup feature from \"Never auto-merge\" here would build the rejected policy. The human-judgment worry survives only as the bar and the gate.\nstrength: strong\n\n### F2\ntarget: docs/research/unbuilt-features-research-mar31.md:24\ntarget_words: \"Two vote buttons per claim: well-argued / poorly-argued (left), agree / disagree (right)\"\nolder: docs/research/session7-deep-review-and-strategic-assessment.md:421\nolder_words: \"Option B — no separate quality vote. The QBAF badge IS the quality signal (computed from CQ evidence). Human vote is ONLY agree/disagree.\"\nkind: contradicts\nwhy: The same Session 7 decided Option B and cites the same LessWrong finding (over 95% of two-axis votes go the same way). TODO.md line 3825 marks single-axis voting DONE on 31 March. The 31 August no-verdict-acts blessing also keeps only the stance vote. This note's Priority 1 recommends the design that was declined that day. Its Bayesian-smoothed agreement percentage is the part that was built.\nstrength: strong\n\n### F3\ntarget: docs/research/walton-argument-schemes.md:329\ntarget_words: \"The number of answered vs. unanswered CQs provides a natural gradient for argument strength\"\nolder: docs/research/session26-open-rulings.md:339\nolder_words: \"It feels rather obvious that the answers should move the link's strength in proportion to how well supported it is.\"\nkind: later-correction\nwhy: Ledger Q3, ruled in direction 2026-09-28. Critical-question answers are weighed by their own standing and never counted. Each carries a direction and an attackable strength tier. Counting survives only as labelled coverage (\"four of six questions addressed\"). This note's gradient from answered versus unanswered CQs, and its INCOMPLETE/VALID states keyed to all-answered, are the counting rule being replaced. Tiers and arithmetic stay open.\nstrength: strong\n\n### F4\ntarget: docs/research/scientific-method-two-steps-and-deliberus.md:184\ntarget_words: \"In topic domain X, decompose claims from worldview groups A and B to depth N. Measure whether the deepest shared premises converge or diverge.\"\nolder: docs/research/philosophical-foundations-stress-test.md:339\nolder_words: \"ceiling as unfinished decomposition cannot lose, and a wager that cannot lose is not a wager.\"\nkind: later-correction\nwhy: The July stress test showed that a convergence test fails to be a test if every ceiling counts as unfinished decomposition. Parity, permissivism, hinge and essential-contestation endpoints must be first-class outcomes that count against convergence. convergence.md adds that asymmetric ingestion can manufacture agreement. The depth-N protocol proposed here has no way to return a ceiling as a loss, so it is not yet falsifiable as written.\nstrength: strong\n\n### F5\ntarget: docs/research/definitions-as-alignment-tests.md:92\ntarget_words: \"A goal-displacement question** attached to any claim that others must recognise something: would\"\nolder: docs/research/session26-open-rulings.md:699\nolder_words: \"The question *say it without the word; does anything still divide you?* is kept as a probe.\"\nkind: already-answers\nwhy: Ledger AC1 to AC4 were blessed by the founder on 2026-10-07 (\"Blessing all proposals in the wanted-words walkthrough\"). A word's standing is recorded beside its senses and for whom. A move on a wanted word is stored as somebody's claim with its kind. Where two people want one word it stays on both senses with their names, and \"who gets the word\" is its own outcome. Herrán's test question is close to that ruled probe. Proposals 1 and 2 here should be reconciled with AC1 to AC3 before anyone rules.\nstrength: strong\n\n### F6\ntarget: docs/research/walton-argument-schemes.md:494\ntarget_words: \"found **97% of arguments in political debate** could be classified into Walton's schemes (491/505 relations).\"\nolder: docs/research/scheme-set-exhaustiveness.md:165\nolder_words: \"**0 of 3,666 scheme-bearing edges**, and 0 of the 3 typed termini\"\nkind: narrows\nwhy: Measured live on 2026-09-11, the shipped classifier had never put `does_not_fit` on any edge, because the prompt did not say when to use it. Forced classification cannot show coverage. Once explained, the value fired on a real no-fit case (about 1 to 3 in 19 pairs). On 2026-09-13 the founder blessed a single exit and retired `default_inference`, the bucket this note's 3% remainder was tagged with. So 97% classifiable is a literature figure, not a coverage claim for this system.\nstrength: strong\n\n### F7\ntarget: docs/research/qbaf-gradual-semantics-research.md:65\ntarget_words: \"Every CQ contributes proportionally. Many weak supports can collectively strengthen; one strong attacker doesn't erase everything\"\nolder: docs/research/session26-open-rulings.md:363\nolder_words: \"one answer on a claim's only edge moves it by under 0.0005\"\nkind: contradicts\nwhy: Ledger Q27, measured 2026-09-27 on the live arithmetic. A one-question edge answered as far as answers go reaches only 0.504 and stays unsettled. Two questions answered in full reach 0.524. QEM's impact is second-order near zero (strength-calculus-literature/bayesian-and-informal-logic.md § 2.7), so lone weak reasons are nearly invisible. The ±6 energy scale and the 0.91 and 0.20 worked examples here are not what the shipped system produces, and the fix is an open decision.\nstrength: strong\n\n### F8\ntarget: docs/research/qbaf-gradual-semantics-research.md:98\ntarget_words: \"All CQs unanswered (0.5 each): energy ≈ 0 → strength = 0.5 (agnostic)\"\nolder: docs/research/strength-layer-audit.md:63\nolder_words: \"summing raw strengths would let three unexamined supporters manufacture a strong claim out of silence\"\nkind: contradicts\nwhy: The note's function sums raw polarity strengths. Unanswered CQs at 0.5 then give zero energy only when supporters and attackers are equal in number (three unanswered supporters give about 0.85). The shipped claim level sums offsets from 0.5 on purpose, so an unexamined edge contributes exactly zero. Anyone re-implementing from this snippet would reintroduce the defect the audit says it avoids.\nstrength: strong\n\n### F9\ntarget: docs/inductive-reasoning-research.md:248\ntarget_words: \"User calibration scores tracked over time (Metaculus-style). Better-calibrated users' estimates weighted more heavily.\"\nolder: docs/research/cognitive-bias-codex-and-human-contribution.md:141\nolder_words: \"proposal to weight contributions by a contributor's demonstrated rationality is\"\nkind: later-correction\nwhy: Item 4 of this page (line 242) already carries a supersession note about weighting by calibration score, but item 7 and the line 189 sentence (weight each user's estimates by calibration) stand uncorrected. The bias note finds no general susceptibility factor and test-retest of 0 to .82, and calls rationality-weighting of contributions measuring noise, \"an empirical kill\". That evidence concerns bias tasks, not forecasts on resolved claims, so the page should say which it relies on.\nstrength: strong\n\n### F10\ntarget: docs/inductive-reasoning-research.md:73\ntarget_words: \"the platform can validate each step's entailment from its premises and flag where reasoning chains become unreliable\"\nolder: docs/research/pass-quality-and-the-choice-of-model.md:153\nolder_words: \"The five-step scale of how far a link bears did not, on any model.\"\nkind: narrows\nwhy: The page promises \"step 3 has only 0.42 entailment probability\". The corpus measured that machine readings of degree are not repeatable. The five-step scale of how far a link bears was not steady on any model, and the linking step asked twice gave the same relation on only 85% of pairs. An unsteady reading cannot be bounded, so the corpus writes no machine number as a reading. A yes-or-no per part was repeatable (45 of 45), but a degree is not.\nstrength: strong\n\n### F11\ntarget: docs/research/scientific-method-two-steps-and-deliberus.md:177\ntarget_words: \"Ratio of user-confirmed vs user-rejected extraction claims (pipeline accuracy)\"\nolder: docs/research/earning-the-right-to-write.md:245\nolder_words: \"If people rubber-stamp, the correction rate looks superb and the pass gets promoted on\"\nkind: later-correction\nwhy: The proposed \"pipeline accuracy\" ratio is the correction record. The trust-ladder note shows it fails silently: if reviewers rubber-stamp, the rate looks superb and a pass is promoted on false evidence, and the failure looks exactly like success. Its remedies are a blind subset so anchoring is measured, and natural errors timed to first objection. Logged as proposed, this metric would read acceptance as accuracy.\nstrength: strong\n\n### F12\ntarget: docs/research/extraction-improvement-research.md:420\ntarget_words: \"Add five values: `observed | causal_hypothesis | prediction | counterfactual | feasibility`. Half-day.\"\nolder: docs/research/dogfood-run-6-israel-palestine-cross-domain.md:203\nolder_words: \"Roughly a third of the claims in both texts are neither observations, nor causal hypotheses, nor predictions\"\nkind: contradicts\nwhy: Run 6, a legal-doctrinal pair, found about a third of claims fit none of these five values, nor normative assertion. They are doctrinal-interpretive: a legal proposition follows from a text. It flagged `taxonomy_strain` and called this the third taxonomy in three runs with the same hole. It asked the founder whether such enums should carry an explicit unclassified value rather than force a nearest fit. \"Fixes empirical overuse\" holds for the first-run register only.\nstrength: strong\n\n### F13\ntarget: docs/research/support-semantics-and-the-conjunction-problem.md:45\ntarget_words: \"where an unsupported claim sits at an agnostic 0.5 rather than being unable to stand\"\nolder: docs/research/relevance-and-the-two-factors.md:271\nolder_words: \"that there's SOME truth value to a claim if it's in the middle then that'd be misguiding if\"\nkind: narrows\nwhy: The note treats the agnostic 0.5 for an unsupported claim as the gentler, current reading. On 2026-09-29 the founder objected that a middle position implies some truth value where there are no good, relevant reasons. Q34 (ruled 2026-09-30, for now) adds a third display state, \"nothing stands (N checked)\". Q35 adds that 0.5 as an indifference prior cannot hold for all claims at once. So the stricter evidential reading is partly adopted at the display, and 0.5 is under challenge.\nstrength: strong\n\n### F14\ntarget: docs/research/support-semantics-and-the-conjunction-problem.md:83\ntarget_words: \"**no argument may both support and attack the same target**\"\nolder: docs/research/extraction-improvement-research.md:306\nolder_words: \"No claim can both SUPPORTS and ATTACKS the same target claim\"\nkind: same-idea-earlier\nwhy: The note presents this as a ready-made rule supplied by the 2026 aggregative-semantics paper and \"the first concrete constraint\" for the ontology layer. The 28 March extraction research already lists it as a rule-based Pass 4c graph-consistency check needing no LLM. It sits beside a DAG check on DECOMPOSES_INTO and a value_premise-should-not-support-empirical check. The rule predates the paper, and a design for enforcing it already exists to build on.\nstrength: medium\n\n### F15\ntarget: docs/research/perceiving-possibility-and-the-analogy-organ.md:128\ntarget_words: \"would tell daemons **where to roam** (high temperature: propose, break, re-frame; low\"\nolder: docs/research/major-transitions-heredity-and-evolvability.md:605\nolder_words: \"Roaming by temperature should be capped by the region's review capacity.\"\nkind: narrows\nwhy: The proposed region-level temperature already has a stated guard. Copycat's temperature is an exploration rate, and the error-threshold argument says variation must not outrun what selection can purge. That note was revised the same day as ledger T2: what must stay under a tolerance is the daemon's unexamined error, not its proposal count. Review capacity caps the rate only while the error rate is unmeasured, which for the analogy daemon may never lift. T2 is open, recommended adopt. Whoever rules on temperature should rule on T2 too.\nstrength: medium\n\n### F16\ntarget: docs/research/extraction-improvement-research.md:273\ntarget_words: \"Take 2/3 consensus on: claim text, claim type, epistemic status\"\nolder: docs/research/writing-out-the-unwritten/outside-reading-dependable-small-tasks.md:354\nolder_words: \"Voting over repeated samples works only with independent errors and exact-match outputs.\"\nkind: narrows\nwhy: This note proposes two-thirds voting over reruns of one model on free-text claim wording. The later reading finds voting works only with independent errors and exact-match outputs. A panel of nine judges from seven model families carried only about two independent votes. Reruns of one model share errors, and claim text is not exact-match. Agreement would overstate reliability. Vote on closed labels only, and measure same-answer-twice steadiness first.\nstrength: medium\n\n### F17\ntarget: docs/research/listening-for-the-calculus-and-sturdy-microtasks.md:113\ntarget_words: \"survive small errors in the probabilities of their parts. The same property is what a strength\"\nolder: docs/research/qbaf-gradual-semantics-research.md:155\nolder_words: \"clamping a single child granted↔denied swings ±0.15 energy through the shallow region of h(x), moving the parent ~0.022 in strength\"\nkind: supplies-a-missing-piece\nwhy: The strength layer already has a measure of the sensitivity Irving's condition asks for. At shipped weights, flipping one child moves its parent about 0.022, roughly two percentage points, and the note calls the hinge a ranking instrument, not a drama meter. A wrong machine reading of one part therefore moves a claim little today. The condition can be tested against that figure instead of argued, and the subclaim weight is the knob that would change it.\nstrength: medium\n\n### F18\ntarget: docs/inductive-reasoning-research.md:246\ntarget_words: \"if X were removed, the overall argument strength drops from 0.82 to 0.31\"\nolder: docs/research/qbaf-gradual-semantics-research.md:155\nolder_words: \"The hinge is therefore a *ranking* instrument at current graph maturity, not a drama meter\"\nkind: narrows\nwhy: The page's counterfactual example shows a drop from 0.82 to 0.31. The July production calibration at shipped weights found a single leaf flip moves its parent about 0.022, so a counterfactual removal reads as a modest ranking signal. It is not a dramatic fall. The sentence reads as a promise the shipped calculus does not keep. Its note says single cruxes would swing conclusions only if the subclaim weight is retuned from observed engagement.\nstrength: medium\n\nNoticed: Other leads I read and did not report. (a) walton-argument-schemes.md:461-464 (\"Limit recursion depth (e.g., 2-3 levels)\", \"AI-assisted assessment at deeper levels\") against ledger T1, session26-open-rulings.md:471, founder \"Yes.\" 2026-09-29: depth is capped by a measured error rate, and the Session 7 spec decided CQ recursion depth is \"Flywheel-gated\" (session7-deep-review-and-strategic-assessment.md:363). (b) scientific-method-two-steps-and-deliberus.md:44 and :50 (\"green badge = both steps done\") against bengio-safety-from-honesty-and-deliberus.md:70 (\"A claim can be strong because nobody attacked it yet\"). (c) the-irit-harvest.md:31 (cite the high-attribution edges) against strength-calculus-literature/computational-argumentation.md:360 (a contribution mixes both factors and cannot serve as a relevance input without circularity). I dropped it because \"computational-argumentation\" appears in the harvest as the phrase \"computational-argumentation group\", which might trip a name check. (d) the-irit-harvest.md:53 is confirmed on its textual side by auto-connect-upgrade.md:7 (claim-level cosine: 402 within-debate candidates, none at the link threshold). (e) definitions-as-alignment-tests.md:88 (\"split it into the criteria\") meets the 2026-09-21 canon line in a-cut-through-a-concept-is-an-argument.md:11 (a cut through a concept is an argument, not a split). A stipulated legal definition passes the recomposition test and the popular sense does not. (f) scientific-method-two-steps-and-deliberus.md:144 (revealing a difference changes the communities) is stronger than does-descending-change-the-descender.md, which finds the effect appears, disappears and reverses, and calls it not currently checkable. (g) inductive-reasoning-research.md:242 points to what-human-judgment-is-for.md § 4c for the empirical kill, but the text sits in § 4c-ii (cognitive-bias-codex-and-human-contribution.md:141 is the cleaner home). (h) The hook notes on several files were treated as data and not acted on. I wrote no files.\n"}