{"path":"research/what-a-rule-rests-on/older-corpus-bearings-strength-and-machine-work-2026-10-08.md","content":"# What the older corpus holds on strength, incentives and machine work, set against the thirteenth walkthrough (2026-10-08)\n\nOne of three read-only readers (Sonnet) sent late on 2026-10-08, after the founder asked: *\"Is there\nany part of the corpus that bears on any aspect or even any single sentence of the latest audio\nwalkthrough that you haven't fully grasped or related it to?\"* This reader took parts one, five, six\nand nine of the walkthrough and its closing. The leads it was given were Claude's guesses, marked as\nsuch. Its report is below, unchanged. Claude's reading of the three reports is the note's § 14\n([../what-a-rule-rests-on.md](../what-a-rule-rests-on.md)).\n\nIn the report, \"the four notes\" are philosophy-of-the-unwritten-step.md, writing-out-the-unwritten.md,\nthe-squirrel-and-the-strength-calculus.md and what-a-rule-rests-on.md. A bare number in the third\ncolumn, or \"W:\" with a number, is a line of the kept walkthrough text.\n\n---\n\n**Paths.** R = docs/research/ · W = docs/listening/2026-10-08-eight-proposals-pondered-again.md · C = the repository root · ledger = R/session26-open-rulings.md. I read the five W parts, the four notes, every lead section and the code named below. I grepped the four notes and W for each lead. Nothing was written.\n\n| Lead | Corpus passage | W sentence | Already related? | What it would change |\n|---|---|---|---|---|\n| L1 | R/incentives-analysis.md:116 \"A rational actor is legible exactly where legibility helps and vague everywhere else.\" | 161 \"Showing your work can only cost you.\" | No. No hit for incentive, legibility or vagueness. | Names a second cost, exposure (:17), beside your arithmetic one. Supports are self-serving and attacks altruistic (:79), so expect planks over-supplied and cases against under-supplied. Its defence (:120, opponent proposes, you contest) is AD24/27/28 renamed, and needs the machine to write for every link. Lead half right: \"vagueness\" is a contributor's choice; here the machine imposes the exposure. |\n| L2 | R/weakest-link-arithmetic…:352 \"three parts each established at 0.9 lift their parent only to 0.507\"; :670 \"adding an irrelevant part… takes any claim out of the green band\" | 157 \"the code rewards you with a hair\" | Partly. Whisper, 0.507 and the shared gate: yes (squirrel:130-138; weakest-link:369). 9e and the TODO items: no. | Lead wrong on the whisper. Your closing question is TODO.md:1822 (stack item 3), with a lean on record. TODO.md:1846-49 already found \"the core contribution act would lower strength by default\" and answered with weights 0.5/0.25/0. 9e's 0.75 is moot while required parts hold the claim near 0.511, and goes live when you rule lift yes (9d:591: a lift raises \"the more a spurious part can take away\"). |\n| L3 | R/derivability…:176 \"selective scrutiny stops paying, because the scrutiny was already there for everyone\"; :197 the lift measures the share of the derivable requirement space established | 569 \"any step can be written out, the moment somebody asks\" | No | The closing asks about soundness; the corpus's variable is coverage. A sentence-level check found 2 of 47 (:137), yet squirrel:134 makes the blind reader the gate. Ledger AD29's \"first level eagerly\" is absent from the spoken text; eager and uniform is what removes the payoff for selective asking. A rule derivable for any link is a standard question, weight 0.25 (C/deliberus/support_semantics.py:124), not a full cap. |\n| L4 | R/earning-the-right…:371 \"The splitter may write what it FOUND and not what it ADDED.\"; R/microtasks…:252 \"adds nothing the claim does not contain, because a written part is attributed to the claim's author\" | 457 \"Committed by the step. Not asserted.\" | Partly. AD18 \"found before made\" (writing-out:510) draws the line, uncited. | 13 of 15 rules stand in neither text (writing-out:452): added, so propose-only. The more-places-to-disagree gate (:38) admits a pass to the ladder; it does not set the rung (:72). The 2026-09-21 ruling covers found parts; squirrel:40 stretches it to rules. Code: an unblessed pass presumes nothing (support_semantics.py:205). |\n| L5 | R/proving-a-pass-reliable…:96 \"With no measured rate, it may not build on machine output at all\" | 593 \"Neither direction has an end.\" | Partly. AD19 and writing-out:516 cite T1; Parts 9 and 10 do not. | The stack is machine link (no labelled extraction error rate, :85), machine rule, machine counterexample, machine reader. Put a depth cap in Part 9. §1 also says review need not keep pace with volume, so AD28's person-gate per rule rations width; count depth. |\n| L6 | R/store-the-uses…:129 \"A reading is not stored as a fact. A judgment is not stored without its record.\"; R/storing-the-record…:125 | 149 \"The sentences we write out under a link are the planks of that bridge.\" | No. The record appears once, as crash recovery (writing-out:475). | The wording is a recorded judgment, one pointer on the link. AD2/AD6 \"stored as text\" is what the law retires. \"Committed by the step\", a rule's count and AD24's \"complete for a named reader\" are readings or record queries. \"On request\" should ask the record first (Z12): open writing \"comes back in different words\" (writing-out:286). |\n| L7 | R/latent-scaffolding…:71 \"could this node be re-rendered from code plus references alone?\" | 165 \"The place under every link exists.\" | Partly. writing-out:269 says \"the way a critical question does\"; the difference is undrawn. | Not the same shape. A question is template plus slots; a rule is varying prose, and a wrong slot \"renders a confident nonsense sentence\" (:127). So a potential must first be a stored judgment. No question may appear without both endpoints (:137), which bars listing a rule \"stored once with its count\" (AD18) apart from its links. |\n| L8 | R/the-ratification-record…:185 \"The graph has contributions and telemetry, nothing else.\" | 511 \"The rule under your step is yours when you say so.\" | No | Mostly fits: own, replace and withdraw are sole-mintership contributions (:189-196). It collides only with a stored third standing. That doc dissolves `confirmed` into provenance plus inferred telemetry (:231) and records a click as a click (:52), so \"I pay\" cannot read as examined. Make \"committed\" a displayed reading; that shrinks the ontology call. Assent to a machine's sentence is telemetry (:129). Lead partly wrong. |\n| L9 | R/graph-daemons…:39 \"the two signals must differ in mechanism or in model family, not merely in invocation\" | 403 \"Every rule a machine writes is shown together with the strongest case against it.\" | Partly. writing-out:537 and philosophy:374 hold the same-family point; co-stimulation and \"homework\" (analogy-daemon:50) are uncited. | Lead right. Count a counterexample only if looked up in a text (Van den Haag was) or from a second family. The writer's family imagining one is one detector twice. AD27's \"met none\" proves little when the same hand looked. |\n| L10 | R/soft-canonical…:378 case 2 \"one evidence node reaching one aggregation point via several paths\"; C/deliberus/evidence_routing.py:125; R/retrieval-instruments:95 | 375 \"It never asks whether two strands are one strand, counted twice.\" | Partly. squirrel:56 and what-a-rule:387 point to Q22/Q4. The taxonomy and the built fair-share are uncited. | Two sufficient sets sharing a plank are case 2. Fair-share is built, unadopted, and only lowers. Its docstring (:215) says sharing cannot inflate under weakest-part, which backs your chain/cable split. A written-out step is the scheme-instance ID that Yu's test lacks (:95). W:379 \"Any two would do\" is not the `alternative` setting, where one suffices (support_semantics.py:17). |\n| L11 | R/convergence-wager-red-team…:89 \"a large argument containing a fatal error that neither the honest opponent nor the judge can localize\" | 367 \"Inside one set of conditions, every part is needed.\" | No. No hit for obfuscated, Barnes or stability. | Half right. Pruning does not answer it: each thin plank passes the blind reader locally, and min is count-blind (weakest-link:184). CLAUDE.md:695 names stability under re-decomposition as the mitigation. AD26's 12 of 24 links with two sufficient sets (philosophy:302) is that measurement, read as \"several arguments\"; the corpus would also read it as a warning. Say both. |\n\n**The three that matter most**\n\n1. **\"Can only cost you\" fails for the links you would write out first.** A link a machine guessed presumes relevance 0 (C/deliberus/relevance.py:24, 53), and an examined relevance claim replaces the presumption whatever its value (relevance.py:84; C/deliberus/graph/badge_inputs.py:399-404). The 871 links of AD8 and AD13 are that kind. On them, writing out and examining can only raise a link; even a refuted claim at 0.2 beats 0. Squirrel:70 scopes its 0.519 to \"an extracted link\"; Part 1 and the closing drop the scope. Planks a machine wrote also start at weight 0 (support_semantics.py:205), so \"sound lifts, rotten costs\" needs a person to have examined each plank. That is the rare case.\n2. **The closing question measures the wrong variable.** The open call is coverage of the derivable requirement space (TODO.md:1822-24, 1911), not whether the parts are sound. A same-family blind reader that \"takes for granted what the writer takes for granted\" (writing-out:437) cannot be the gate, and frame lock-in is invisible to any sentence-level check (TODO.md:1886). Ruling yes opens the 9e door.\n3. **One hand all the way down.** The corpus lets a machine write what it found, build only as deep as measured, and act on two independent detectors. The spoken design has a machine-drawn link, a machine rule, a machine counterexample and a machine completeness check. Fix: counterexamples and cases found by lookup, and rules propose-only.\n\n**Contradictions with the older corpus**\n- Base relevance 0 versus \"can only cost\" (W:157-161, 703; AD22's \"earns its author almost nothing\").\n- AD22 and squirrel \"required parts\". The 2026-09-21 ruling is about found splits. AD17 says a machine sentence \"weighs nothing\". Code gives it weight 0. A part the claim does not contain is \"invented\" and presumes nothing (support_semantics.py:146-151).\n- W:569, 597 \"when somebody asks\" drops \"with a reason\". Philosophy:139 and AD22 keep it. what-counts-as-evidence.md:224 says a bare demand becomes \"a stalling weapon\" (ledger Q2, :338). AD29 inherits the omission.\n- W:367-375 cable. Pollock, source of the chain rule, \"rejects accrual outright\" (weakest-link:56-58). The corroborative channel sides against him on purpose. Adding across sets is the contested choice, not just its double counting. Amgoud and Ben-Naim (:104) say no arithmetic is neutral.\n- W:531 and 561 still say agreement on cases comes first, which your header (81-85) and AD28 withdraw.\n- AD28's stored third standing against ratification:231.\n- AD7, \"no machine reading of how far it reaches may weigh a link\", against the ruled Q33 (relevance_judge.py:5), where a second-family machine proposes relevance. See bearing 1.\n\n**Other bearings**\n1. **The machine relevance judge** (C/deliberus/relevance_judge.py; Q33; O13). It is inert, with `--write` awaiting you. Trial: 6 strong, 4 moderate, 2 weak (R/relevance-and-the-two-factors.md:161). No source for the walkthrough or the four notes mentions it. W:643 says the bridge's number \"is where people disagree\" without saying a judge already proposes it. AD22's written grounds are a second route to the same number.\n2. **Your 2026-09-19 reception** (weakest-link:633): relevance \"judged provisionally by the machine/LLM-passes\" and humans challenge. That is closer to AD22 than AD17's \"weighs nothing\".\n3. **Red team Attack C** (:138, :63): narrative on-ramps and score-free nursery. W:507 \"Almost everyone can name their case\" is that on-ramp. The red team also warns (:176) that a descent can \"detonate a functioning fudge for the price of one URL\". \"On request\" is that URL.\n4. **Analogy daemon prior pass** (the-analogy-daemon.md:44) proposes a number. W:651 \"never hands you a number\" drops it without saying so.\n5. **Minor.** \"Two months apart\" (W:435) is 38 days (2026-08-30 to 2026-10-07).\n"}