{"path":"research/corpus-graph/readers/2026-10-27-batch-3-purposes-signposts-and-trials-as-a-reader-returned-it.md","content":"# What a reader found to bear on ten notes on what Deliberus is for, signposts, flywheels, the what-if view, normic rules, cases, citations and the kind of mistake we make (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 10 notes:\n\n- `docs/research/what-deliberus-is-for.md`\n- `docs/research/signposts-for-the-whole-project.md`\n- `docs/research/flywheels-in-building-deliberus.md`\n- `docs/research/the-kind-of-mistake-we-make.md`\n- `docs/research/citations-in-the-passes.md`\n- `docs/research/what-if-and-lenses.md`\n- `docs/research/normic-rules-normative-reasons-and-trying-rules-out.md`\n- `docs/research/what-a-case-counts-for.md`\n- `docs/research/reasons-and-cases.md`\n- `docs/research/how-independent-reasons-add.md`\n\nIt made 105 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\nI read all 10 of the 10 notes in full. All quotes below were checked by script as single-line substrings of the files. None of these pairs is in found.jsonl; 8 of the 10 notes appear there in no pair at all.\n\n### F1\ntarget: docs/research/citations-in-the-passes.md:94\ntarget_words: \"the same two counts on two more sources of other kinds (a news report, a policy paper), against\"\nolder: docs/research/session26-open-rulings.md:728\nolder_words: \"A pass is read by a person before it is proven, and a rate is proven only for the kinds of fault that reading found.\"\nkind: narrows\nwhy: Y9 is ruled and resident. A pass is read by a person, one case at a time, before any rate is measured. The citations plan takes a second and third source and counts reach and duplicates against the session's own hand readings. That is the step TODO.md:3611 records as breaking Y9 in the cheaper-model work, and pass-quality § 10e withdrew it. The first step should be his reading of a few cited claim lists, one at a time.\nstrength: strong\n\n### F2\ntarget: docs/research/what-if-and-lenses.md:346\ntarget_words: \"5. **The arithmetic before the view.** Build Q38 (the weakest premise within an argument; the best argument\"\nolder: docs/research/session26-open-rulings.md:462\nolder_words: \"Within one argument the case facts combine as a conjunction: the product where their reasons share no root, a shared root counted once\"\nkind: later-correction\nwhy: Q47 was ruled 2026-10-25, and the resident line \"within one argument the weakest premise governs\" was reworded to the conjunction on 2026-10-26. Q47 also answers § 9j without a keeper reading, so proposal 5 is half ruled and half overtaken. The § 2 figures, where denying one required part caps the whole \"by the weakest-part rule\", were measured under the rule Q47 replaced. A person building the view should not wait on Q38 as worded, and should rerun § 2.\nstrength: strong\n\n### F3\ntarget: docs/research/normic-rules-normative-reasons-and-trying-rules-out.md:203\ntarget_words: \"the weakest premise within an argument), with no new arithmetic. The hinge score, which already asks how far a\"\nolder: docs/research/session26-open-rulings.md:463\nolder_words: \"Where the dependence between the facts of one argument, or between separate arguments, is not recorded, the combination shows its range\"\nkind: later-correction\nwhy: § 4c says a consequence trial's standing comes from Q38, the weakest premise, \"with no new arithmetic\". Q47 (ruled 2026-10-25) replaced that with a product of the facts, a shared root counted once. Q48 (ruled 2026-10-26) says an unrecorded dependence shows a range. So a trial's conditional output, \"if these scoped rules hold\", is a range plus a labelled estimate. The trial's output form is not settled by Q38, and the hinge ranking sits on the new arithmetic.\nstrength: strong\n\n### F4\ntarget: docs/research/what-a-case-counts-for.md:321\ntarget_words: \"6. **AY6, a one-case claim is read through the narrowest rule for which a suitable count exists**, a\"\nolder: docs/research/session26-open-rulings.md:1055\nolder_words: \"Several statistical generalizations for one case are never handed to an engine together: the narrowest counted reference class answers alone\"\nkind: narrows\nwhy: AM18 was ruled 2026-10-24 and built, with a test. The narrowest counted reference class answers alone. Otherwise rules combine by the graph's own rule: log-odds only where the pair was read independent, else the best, and a pair read as one story owes its shared cause as a claim between. AY6 adds Kyburg's condition that a narrower class overrides only where its interval disagrees. AM18 has no such condition and cites a measured failure (an engine answered 0.732, above both rules). Building AY6 must reconcile the two.\nstrength: strong\n\n### F5\ntarget: docs/research/what-deliberus-is-for.md:139\ntarget_words: \"I'm not gonna let that stand in the way of coming up with or noting down ideas that are relevant to Deliberus.\"\nolder: docs/research/flywheels-in-building-deliberus/checkins.md:17\nolder_words: \"a week with more rows added than ruled makes the next session hold new rows and offer a ruling sitting\"\nkind: contradicts\nwhy: Written after his bound of 2026-10-11, the Monday check-in offers him a trigger on the open-rows signpost. When a week adds more rows than it rules, the next session holds new rows. His bound makes the count of open positions something a session says, never something that stands in the way of an idea or of writing one down. Anyone ruling AH2's second step or that trigger should see that option (a) is the hold his bound excluded. His reply is only predicted (listening 2026-10-23).\nstrength: strong\n\n### F6\ntarget: docs/research/what-deliberus-is-for.md:133\ntarget_words: \"None of this is built, and the first is a habit, not a mechanism. A habit kept by a session that\"\nolder: docs/research/what-agent-memory-research-says-about-agent-surfaces.md:20\nolder_words: \"a filled memory store made no memory call in 114 turns, while the same items delivered by its\"\nkind: same-mechanism-elsewhere\nwhy: The target rates its session habit weak and proposes a block in the always-loaded file. A measurement in a different field supports it. An agent that owns a store mostly does not consult it (no call in 114 turns), while the harness delivering the same items at set moments reached 138 of 138. The note adds the project's own instance, a session that opened the ruling 49 minutes before breaking it. So AH2's \"say it at the start\" needs a delivery mechanism, not only AH3's placement.\nstrength: strong\n\n### F7\ntarget: docs/research/what-deliberus-is-for.md:82\ntarget_words: \"I believe the vast majority of people have good hearts.\"\nolder: docs/research/session26-open-rulings.md:1400\nolder_words: \"Most disagreements, decomposed, show shared care held with different beliefs, weights and circles of concern\"\nkind: already-answers\nwhy: The target builds purpose 2 on this vision-page sentence, and purpose 3 on \"whichever way it comes out\". BD28 (ruled 2026-10-25, resident) already narrows the belief. Decomposed disagreements show shared care held with different beliefs and weights. The commonest error is thinking the other side's heart is bad. A descent that finds no shared care counts against the wager. The word \"just\" did not survive. Purposes 2 and 3 should use this ruled wording and its falsifier, and not re-compress the belief.\nstrength: strong\n\n### F8\ntarget: docs/research/flywheels-in-building-deliberus.md:198\ntarget_words: \"It does not lower the draw on his attention; it adds hours in which his\"\nolder: docs/listening/decisions-register.md:22\nolder_words: \"43 decisions; 26 open, 14 settled by their row, the rest duplicates.\"\nkind: supplies-a-missing-piece\nwhy: Flywheel 9 says listening only adds hours of attention. The register of 2026-10-25 counts the questions walkthroughs put to him, and 26 of 43 are still open. Listening therefore feeds his ruling, which § 3 names as the constraint, and it builds the queue Goldratt's \"subordinate\" step warns against. Its gearing is positive for supply but negative for the constraint, and the note currently gives only the first. Listening's output should be weighed beside the open-rows trigger.\nstrength: medium\n\n### F9\ntarget: docs/research/what-if-and-lenses.md:123\ntarget_words: \"- **Do not sell the view as humbling or moderating.** Explaining how a policy would work lowers people's\"\nolder: docs/research/does-descending-change-the-descender.md:121\nolder_words: \"Whatever epistemic humility descent produces, it is produced\"\nkind: same-idea-earlier\nwhy: This note read the same literature on 2026-08-27 and found more. The effect appears, disappears and in one study reverses. It needs self-generation, since reading another's explanation gives none. Decomposing a value claim into reasons is the condition that does not moderate and may entrench. So a two-person lens read passively cannot claim humility for the reader, and the benefit belongs to the person who supposes premises themselves. Proposal 8's measure and the first workshop should test that split.\nstrength: medium\n\n### F10\ntarget: docs/research/normic-rules-normative-reasons-and-trying-rules-out.md:142\ntarget_words: \"It needs no language model once the rules and facts are written\"\nolder: docs/research/tools-for-checking-inferences.md:30\nolder_words: \"language, and that translation is where it fails: a verifier that has the model write first-order logic\"\nkind: supplies-a-missing-piece\nwhy: The deterministic trial begins after \"written as claims\", and that step is where the measured failures sit. A verifier that has the model write first-order logic could check only 15 to 45 percent of chains. A model writing ProbLog unaided gave no valid answer in about three cases of four. The note's own design answer is that the model fills slots in its own words and plain code builds the engine input. That is what makes § 4a's \"no model\" safe, and AX5 builds on it.\nstrength: medium\n\n### F11\ntarget: docs/research/citations-in-the-passes.md:41\ntarget_words: \"Today every pass that points into the source asks the model to copy a passage into a field\"\nolder: docs/research/store-the-uses-read-the-meaning.md:202\nolder_words: \"The stored quote is in the source character for character, in exactly one place\"\nkind: supplies-a-missing-piece\nwhy: This measures the live baseline on 1,144 claims: 81 percent exact, 94 percent after typography is evened out. Citations is measured only on one article (GLM 57 percent, Gemini 97 percent), so the gain over what is stored today is about 6 to 19 points. The same table finds a claim's words are only 34 percent from its passage, 29 percent from nowhere in the source, and 15 percent of claims share a passage. A cited passage anchors a claim but does not replace the rewritten sentence.\nstrength: medium\n\n### F12\ntarget: docs/research/the-kind-of-mistake-we-make.md:97\ntarget_words: \"the notes; a principle that is not in view at the moment of generation loses to the template.\"\nolder: docs/research/what-agent-memory-research-says-about-agent-surfaces.md:20\nolder_words: \"a filled memory store made no memory call in 114 turns, while the same items delivered by its\"\nkind: same-mechanism-elsewhere\nwhy: § 6.1 says a principle not in view loses and names a remedy, the corpus graph asked before writing. The memory research measures that mechanism. An agent that owns a store does not consult it (0 calls in 114 turns), while harness delivery at set moments arrived 138 of 138. So the remedy works only if the lines are handed over at the act (AK3's reply-time search), not left to be looked up. That also makes the \"test is a law\" point in § 6.3 stronger.\nstrength: medium\n\n### F13\ntarget: docs/research/the-kind-of-mistake-we-make.md:399\ntarget_words: \"The always-loaded files measured that night: 514,934 bytes for this project's `CLAUDE.md`, 391,792 for the\"\nolder: docs/research/flywheels-in-building-deliberus/checkins.md:25\nolder_words: \"It grew 70,211 → 78,128 words (+11 %) in four days while no line was retired\"\nkind: supplies-a-missing-piece\nwhy: § 10 adds BD11, BD13 and BD15 as resident lines, and § 11 proposes further global rules, but it measures size only once. The Monday check-in of 2026-10-26 finds the resident file grew 11 percent in four days with nothing retired. It names the file, and no longer his ruling, as the constraint, and puts a word cap to him. A further resident-line candidate should now say what it displaces.\nstrength: medium\n\n### F14\ntarget: docs/research/signposts-for-the-whole-project.md:150\ntarget_words: \"Legibility that wins is not turned into control (Porter's dark rider)\"\nolder: docs/research/legibility-under-power-red-team.md:219\nolder_words: \"The Legibility Rule, applied reflexively, indicts its own home\"\nkind: supplies-a-missing-piece\nwhy: The row's sign is \"requests to remove or promote; a written constitution\". That sees only outside pressure. The July red-team names the operator's own constitutive moves (prompts, model choices, scheme taxonomy, semantics selection, curation) as challengeable in-graph by no one, and a solo operator as a single point of capture. frontier.md calls this the operator-reflexivity gap. Add a sign for the operator's own moves (how many are claims with roles), or the row will read green while that gap stands.\nstrength: medium\n\n### F15\ntarget: docs/research/signposts-for-the-whole-project.md:291\ntarget_words: \"- **Twenty-five is too many to watch.** The planning literature watches only what is both load-bearing and\"\nolder: docs/research/drowning-in-claims.md:83\nolder_words: \"Candidate early-warning dials, each countable from the graph\"\nkind: supplies-a-missing-piece\nwhy: The register has no row for claims minted faster than they are examined, though it is the failure it calls \"padding\" and \"saturation\". An earlier note already lists six dials countable today: duplicate rate at mint, orphan rate, debt ratio, unclassified rate, scaffolding fraction and the Good-Turing novelty rate. When the register is ranked down from twenty-five, this gives one row with a ready instrument. The debt-ratio dial is tied to the proven bound.\nstrength: medium\n\n### F16\ntarget: docs/research/how-independent-reasons-add.md:507\ntarget_words: \"from the prior to certainty and applies Fréchet's lower bound to their conjunction.\"\nolder: docs/research/theorists-on-normic-rules-exceptions-numbers-and-many-minds.md:142\nolder_words: \"So *the weakest premise governs* is the **upper** Fréchet bound\"\nkind: same-idea-earlier\nwhy: Q48's range, found by Codex on 2026-10-26, is already in a note of 2026-10-22. The weakest premise is the upper Fréchet bound, the product sits below it, and min is the limit of total positive dependence. That note says the ruling and the § 9j proposal differ only in which end of the bounds is the default, and that this is the sentence to put to him. It also cites Pearl: independence is a graph question, and Q40 reads d-separation. This is from memory and unchecked.\nstrength: medium\n\n### F17\ntarget: docs/research/reasons-and-cases.md:281\ntarget_words: \"2. **BA2, a unit's cases are counted under the unit's own rate.** When several cases come from one unit (a\"\nolder: docs/research/saying-what-kind-a-claim-is.md:193\nolder_words: \"the pooling strength is estimated from how much sibling classes differ and never chosen\"\nkind: supplies-a-missing-piece\nwhy: BA2 needs the spread between units, and § 5.8 (Efron and Morris) shows pooled counts beat each unit's own. This note has built that spread: counted bundles under a class, pooling strength estimated from sibling differences, a lone class standing on a flat prior. It was tested on three real tables (the 2015 replication cells: narrow class error 0.059, broad 0.165, fixed pooling 0.080). Two gaps remain: a member's own case is not yet left out of its class's count, and the stored table is not yet on the live graph. BA2's arithmetic is mostly there.\nstrength: medium\n\n### F18\ntarget: docs/research/flywheels-in-building-deliberus.md:167\ntarget_words: \"labelled set is paid for once and admits or refuses every later model in minutes\"\nolder: docs/research/session26-open-rulings.md:728\nolder_words: \"A pass is read by a person before it is proven, and a rate is proven only for the kinds of fault that reading found.\"\nkind: narrows\nwhy: Y9 is ruled and resident. A rate is proven only for the kinds of fault his reading has found, and each fault changes the instructions, so proving first is paid for twice. The answer key therefore admits or refuses a model only on kinds of fault already read. A new kind of fault needs his reading before any rule or bound, so \"only the residue reaches him\" understates his draw. Gearing 5→6 (a proven pass may run unattended) is scoped to kinds read.\nstrength: medium\n\nNoticed: Candidates cut for space. what-deliberus-is-for.md:84 says \"No person has yet read one\"; Y9's gist (ledger row 728, \"after the keeper stickprov\", 2026-10-19), the nine-pairs listening piece (2026-10-18) and how-independent-reasons-add/his-readings.md (2026-10-23) show he has since read machine verdicts one case at a time, so that \"today\" cell is stale. flywheels-in-building-deliberus/checkins.md:23 updates the signposts § 3g reading (106 rows added, 98 ruled, open 424, trigger still unset). what-if-and-lenses.md:358 (success measured by holders) matches incentives-analysis.md:166, where Community Notes is the production version of that test with a reward. Not reported: reasons-and-cases.md:242 and what-counts-as-evidence/grading-evidence-in-science-and-policy.md:854, whose publication-bias search claims fit BA1's shared route. Not reported either: what-if-and-lenses.md:103 (a live link with inverted polarity, 2026-10-22) against the-kind-of-mistake-we-make.md:119 (\"polarity muddle has not recurred\"), since the two are different polarity faults.\n"}