{"path":"research/corpus-graph/readers/2026-10-27-batch-5-accrual-negation-tools-and-the-group-as-a-reader-returned-it.md","content":"# What a reader found to bear on eleven notes on accrual, negation, inference-checking tools, the 2011-2017 group and the yellow link (2026-10-27)\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/tools-for-checking-inferences.md`\n- `docs/research/one-inference-per-link.md`\n- `docs/research/resonance-db-and-negation.md`\n- `docs/research/claim-parts-and-inferential-reasons-2026-10-14.md`\n- `docs/research/pollock-against-accrual-and-the-overlap-check.md`\n- `docs/research/the-yellow-link.md`\n- `docs/research/derivation-sciencepedia-and-the-logic-core.md`\n- `docs/research/what-the-group-held-2011-2017.md`\n- `docs/research/what-he-held-in-the-group-2011-2013.md`\n- `docs/research/deliberus-as-a-desktop-app.md`\n- `docs/research/weighing-answers-and-the-badge-threshold.md`\n\nIt made 112 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 the 11 notes read in full.\n\n### F1\ntarget: docs/research/pollock-against-accrual-and-the-overlap-check.md:217\ntarget_words: \"The best reason alone is the payoff-removal form for reasons, as\"\nolder: docs/research/the-number-in-the-world-and-the-repertoire.md:166\nolder_words: \"The payoff that was removed was the one a reason-giver had for adding a good reason.\"\nkind: narrows\nwhy: The next day (15 Oct) he said that removing the payoff for adding is a mindless intervention and should be lower priority than reflecting on what is going on (same note, line 11). That note ranks payoff removal as the last rung. It shows that removal applied before the objects are told apart removes the wrong payoff, using the 24 reasons filed as parts under min. The Pollock note rests C2 and its pick (best reason as default) on payoff removal as the law, so § 6 should be read as a last-rung move.\nstrength: strong\n\n### F2\ntarget: docs/research/what-he-held-in-the-group-2011-2013.md:52\ntarget_words: \"about: the agree/disagree vote sits on the claim,\"\nolder: docs/research/corpus-graph/decision-clusters.md:230\nolder_words: \"claim's strength; only the arguments around it do\"\nkind: already-answers\nwhy: Cluster C1 records his ruling of 2026-09-09: your vote never changes a claim's strength, only the arguments around it do. Ledger AS3 (ruled 2026-10-22) adds that a vote only makes the voter a holder of the claim or of its exact negation. So the shipped vote on the claim ranks and weighs nothing. Section 1.1 presents it as the one respect in which what shipped differs from what he was adamant about in 2011. That is mostly not a departure from his stance on voting on claims.\nstrength: strong\n\n### F3\ntarget: docs/research/pollock-against-accrual-and-the-overlap-check.md:331\ntarget_words: \"A lift above the best reason is a claim with the lifted reasons as its required parts.\"\nolder: docs/research/how-independent-reasons-add.md:49\nolder_words: \"independence is read (ledger Q38); independent arguments then add in log-odds, which is the rule Q38 left\"\nkind: later-correction\nwhy: Q46 (ruled 2026-10-22, built 2026-10-23) makes independent arguments add in log-odds from the claim's prior. It retires QEM's squash on the reasons channel. Q37 was settled only through Q38, with the lift's storage as a written claim left open. So the § 3 table row \"every support link adds, then QEM\" and challenge C1 (a lift nobody can attack) describe the code before it was replaced. The lift is now ruled arithmetic gated by the independence reading. Whether it is also a stored claim is the only open part.\nstrength: strong\n\n### F4\ntarget: docs/research/claim-parts-and-inferential-reasons-2026-10-14.md:116\ntarget_words: \"Q22 and AD26 leave important aggregation questions open.\"\nolder: docs/research/how-independent-reasons-add.md:49\nolder_words: \"independence is read (ledger Q38); independent arguments then add in log-odds, which is the rule Q38 left\"\nkind: later-correction\nwhy: The audit says aggregation is unruled and that probability adds routes only under stated dependence. Eight days later Q38 and Q46 were ruled: the weakest premise within one argument, the best argument until independence is read, then log-odds addition. They were checked on five cases with published outcomes. The later note, with its Q48 warning, makes the audit's own dependence caveat the stated limit of the rule. Someone using § 7 and § 13 as a to-do list should treat the arithmetic as decided and the independence reading as the live question.\nstrength: medium\n\n### F5\ntarget: docs/research/resonance-db-and-negation.md:157\ntarget_words: \"what to do about it (a lower threshold for pairs across sources, or candidates fetched by a second route) is not decided.\"\nolder: docs/research/jev-and-the-judgement-layer.md:253\nolder_words: \"A widened candidate set mattered more than a better question.** Letting it see pairs the\"\nkind: already-answers\nwhy: A measurement of 2026-10-04 on a decision model: letting it see pairs the embedding had discarded found real matches, and rewording the question found none. The same section measures that similarity at 0.65 called 19 of 51 different items the same, so a lower bar alone costs precision. That favours the second route, wider candidates read with both texts, over a plain lower cosine bar. It was measured on Jev and status sentences, not on the keeper, so it is a pointer and not a result for these two pairs.\nstrength: medium\n\n### F6\ntarget: docs/research/resonance-db-and-negation.md:173\ntarget_words: \"the graph. Count the shortest chain/imagine what the subgraph between them would look like (or does look\"\nolder: docs/research/folded-routes-and-shared-premises.md:340\nolder_words: \"(ledger AG14's two tests); eleven links may hide thirty inferences or be an artifact of automatic linking.\"\nkind: narrows\nwhy: The folded-routes note lists \"distance in links was treated as inferential distance\" among the mistakes of its own first showcase. Extracted links are not floor-level inferences, so a shortest-path count can be small or large for reasons unrelated to how far apart two takes are. His proposal to count the shortest chain is sound only once the chain is written to the floor (AG14, AM14), as the keeper's slots now do. A count over stored links would measure the linker's artefacts.\nstrength: medium\n\n### F7\ntarget: docs/research/tools-for-checking-inferences.md:129\ntarget_words: \"mornings with falling pressure* replaces both broader rates, and nothing is combined.\"\nolder: docs/research/starting-values-and-the-problem-of-the-prior.md:726\nolder_words: \"limits or special cases, and by Stein's result any rule that picks a single level is beaten by shrinkage\"\nkind: narrows\nwhy: AM18 (ruled) lets the narrowest counted class answer alone, citing Reichenbach. The starting-values note records the competing approach: partial pooling, where a narrow class is pulled toward its parents in proportion to how little data it has. By Stein's result, any rule that picks one level loses to shrinkage in total risk. The ledger's BD4 (child pools toward its parents) is built on that. A narrow class with few counted cases is then a thin basis for answering alone; the tools note does not mention the thin-count case.\nstrength: medium\n\n### F8\ntarget: docs/research/claim-parts-and-inferential-reasons-2026-10-14.md:147\ntarget_words: \"one maximal positive contribution gives 0.600; three required parts at 0.9 give about 0.507\"\nolder: docs/research/what-relevance-is.md:473\nolder_words: \"0.9 and 0.8 under a whole give it 0.504, whether the parts are presumed or argued required; the weakest part is\"\nkind: supplies-a-missing-piece\nwhy: Measured 25 Oct: the weakest part is only a ceiling that the whole never reaches, because parts pass strength up only through the additive energy. So \"parts combine by the weakest\" is wrong of the shipped code, and almost nothing flows. This explains the audit's 0.507 for three parts at 0.9. It also means the audit's long § 8 argument about min (t-norm, probability bound) concerns an operator the code does not run on parts. A reader would stop treating min as the live number.\nstrength: medium\n\n### F9\ntarget: docs/research/what-the-group-held-2011-2017.md:433\ntarget_words: \"three two-cycles and one three-cycle on the live graph, one pass computed.\"\nolder: docs/research/strength-calculus-literature/bayesian-and-informal-logic.md:326\nolder_words: \"A directed cycle in an argument graph reads better as a dependence to model (a common source) or as\"\nkind: supplies-a-missing-piece\nwhy: Item 17 calls the cycle policy unstated, with Kronvall's 2012 prune-to-axioms as the only candidate. The Bayesian review (§ 2.8) holds a second one: a directed cycle is better read as shared dependence or circular support than iterated through, and Wieten et al. break such cycles. It lists \"compute through, or treat as a sign of shared dependence\" as an open question. That matches the group member who held that a circle connotes equivalence. So the ruling on item 17 has two named candidates and a literature behind them.\nstrength: medium\n\n### F10\ntarget: docs/research/derivation-sciencepedia-and-the-logic-core.md:130\ntarget_words: \"worth showing to its holder, and makes the shared intersection, not an average, the honest form of\"\nolder: docs/research/legibility-under-power-red-team.md:58\nolder_words: \"The legal-theory generalization is Sunstein's incompletely theorized agreements: well-functioning\"\nkind: narrows\nwhy: The red-team note argues that agreement often holds because parties read one text differently (Good Friday constructive ambiguity, Sunstein's incompletely theorized agreements). Showing it publicly makes the divergence common knowledge and can break the agreement. A lens intersection (AX7: the claims all lenses accept) can be such a fudge. A neighbouring note adds that maximizing common ground rewards vague platitudes. Whoever rules on AX7 should see that an intersection is honest only if the readings behind each shared claim are also shown.\nstrength: medium\n\n### F11\ntarget: docs/research/the-yellow-link.md:46\ntarget_words: \"costless, friendly-sounding option beside the real ones draws the cases that should have been argued (the rule\"\nolder: docs/research/scheme-set-exhaustiveness.md:165\nolder_words: \"beside a second undifferentiated fallback (`default_inference`, 15 uses).\"\nkind: supplies-a-missing-piece\nwhy: This is the measured case behind the rule the yellow-link note cites only in passing. The scheme classifier's does-not-fit value fired on 0 of 3,666 edges while a costless default sat beside it. The same section treats a zero firing count as an alarm. It supports rejecting a context link. It also cautions against the note's remark that the 80 does_not_fit links show \"the confession value doing its job\". Those were chosen by a map builder, not by the classifier.\nstrength: medium\n\n### F12\ntarget: docs/research/the-yellow-link.md:46\ntarget_words: \"costless, friendly-sounding option beside the real ones draws the cases that should have been argued (the rule\"\nolder: docs/research/session26-open-rulings.md:1361\nolder_words: \"Rival explanations of one fact are answers to one question in the graph, each its own claim, with a *something else* answer always present\"\nkind: narrows\nwhy: BC9, ruled the next day (2026-10-25), requires a free \"something else\" answer beside the real ones on a rival set, kept deliberately. So the yellow-link rule that a free option draws the hard cases is not general. It governs a machine's no-answer on a relation menu, which belongs in a gate. A human-facing rival question needs the open answer. Anyone extending the yellow-link reasoning to other menus (for example AM22's question wording) needs to know where it stops.\nstrength: medium\n\n### F13\ntarget: docs/research/one-inference-per-link.md:117\ntarget_words: \"reason-before-verdict test covers the order in our models; whether the provider honours it was not\"\nolder: docs/research/extraction-improvement-research.md:21\nolder_words: \"Gemini 3 Flash SDK has a known schema-ordering bug that breaks CoT; `instructor` works around it.\"\nkind: supplies-a-missing-piece\nwhy: The note leaves open whether a provider honours key order, after reporting that one provider once sorted keys alphabetically. The March research note records a Gemini 3 Flash SDK schema-ordering bug that breaks reason-before-verdict, with `instructor` as the workaround. The keeper ran on Gemini earlier in this chain (resonance note, free tier and fallbacks). So the check has a concrete target: whether each path the keeper and its fallbacks use preserves field order. The first-time-right numbers may depend on it.\nstrength: medium\n\n### F14\ntarget: docs/research/tools-for-checking-inferences.md:113\ntarget_words: \"(proposed as ledger AM17, *the checker talks back*). AM14 already sends a refused link back once; the\"\nolder: docs/research/extraction-improvement-research.md:380\nolder_words: \"failures trigger targeted retries where the error message is fed back to the model.\"\nkind: same-idea-earlier\nwhy: In March 2026 the project already adopted validate-and-retry with the error fed back, via `instructor` and Pydantic validators. It claimed failures fell from about 5-10% to under 1% in practice, with no source. That was a schema check, not a reasoning check. AM17 (checker talks back, up to three rounds) is the same mechanism lifted to inference form. Its evidence (54.6% to 78.0%) is better, but the project's own earlier figure has never been compared with its production retry record.\nstrength: medium\n\n### F15\ntarget: docs/research/tools-for-checking-inferences.md:166\ntarget_words: \"nothing until examined (AG12), so the saving starts where a person wrote or examined the rule.\"\nolder: docs/research/weighing-answers-and-the-badge-threshold.md:51\nolder_words: \"answers left a claim unexamined, because examination has to start somewhere: an answer's support is a reason, a\"\nkind: same-mechanism-elsewhere\nwhy: The weighing build hit a cold start. An unweighed answer moved nothing, and since answers are where examination begins, no claim could ever be examined. The fix was a lowest-tier default (blessed 2026-10-24). § 8 of the tools note has the same shape: a machine-written rule weighs nothing until examined (AG12), so the saving from reuse waits on a person. Anyone building AM16 should ask whether the dark-layer problem recurs for rules. It would need its own default or a first examiner.\nstrength: medium\n\n### F16\ntarget: docs/research/weighing-answers-and-the-badge-threshold.md:95\ntarget_words: \"One account can still answer ten different questions**, one each. The ruling asked for one per question;\"\nolder: docs/research/session29-the-derivability-arc-and-the-conservative-design.md:218\nolder_words: \"the sybil price from one identity to two, which is nothing. It survives as a *lone-error* guard\"\nkind: narrows\nwhy: Session 29, citing Cheng and Friedman, holds that a symmetric function of source count is not sybilproof. A ceiling raises the attacker's price from one identity to two, so it guards only against lone errors. \"One account counts once per question\" is such a rule, and the note's own table (each answer from its own account, 0.622) shows what many accounts reach. So the open budget question is a sybil question, and the one-per-account rule should not be described as protection against gaming.\nstrength: medium\n\nNoticed: `lean-deliberus-analogies.md` § 9.1 items 2 and 5 bear on the tools and one-inference notes, and neither names that note. They say an engine's findings must come back as claims and links the graph can read, never as a verdict on the engine's say-so. They also treat overlapping reference classes as a typeclass diamond whose repair is to count the intersection as its own class. I left both out as confirming or already ruled (AM18, BD4). CLAUDE.md lines 480-482 record a 15-day silent outage (1-15 April) in which instructor's validation retries could never succeed. It is a useful caution for AM17's retry loop: a refusal the writer cannot fix should not loop silently. The group-note cycle count (6 claims) matches the later folded-routes count of \"six claims in a two-link support loop\", so there is no contradiction there. I found nothing new for the desktop-app note beyond the group note's items 5 and 6 (an overlay on any page, a live companion), which the shell's extras do not cover and which are not built.\n"}