{"path":"research/corpus-graph/readers/2026-10-27-grown-4-unwritten-evidence-confirmed-and-ux-as-a-reader-returned-it.md","content":"# What a reader found to bear on eight grown notes and pages: the unwritten step, the kind of mistake, the synthesis plan, evidence, confirmed, UX principles, cognitive bias, writing out the unwritten (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 8 notes:\n\n- `docs/research/philosophy-of-the-unwritten-step.md`\n- `docs/research/the-kind-of-mistake-we-make.md`\n- `docs/research/synthesis-build-plan.md`\n- `docs/research/what-counts-as-evidence.md`\n- `docs/research/what-confirmed-means.md`\n- `docs/ux-principles.md`\n- `docs/research/cognitive-bias-codex-and-human-contribution.md`\n- `docs/research/writing-out-the-unwritten.md`\n\nIt made 93 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\nRead in full: 8 of the 8 notes listed (the brief says seven but lists eight). All 17 quote pairs below were checked by script against the files at the stated line numbers.\n\n### F1\ntarget: docs/research/philosophy-of-the-unwritten-step.md:490\ntarget_words: \"sentence that would close each. Asked twice, it changed side on 6 of the 25. And a step closed\"\nolder: docs/research/session26-open-rulings.md:1051\nolder_words: \"We need method, not repetition, without clarification, the answer is not liable to change much.\"\nkind: already-answers\nwhy: Row AM14 is RULED 2026-10-26 (founder: \"Rule it and build it!\"). It says the floor of a step is checked in code from slots, \"never by a second reading\". The 13 October paragraph here proposes a second reader that finds a gap and, asked twice, changes side on 6 of 25. The ruling uses that instability as the reason against asking twice. Anyone building the floor check from this paragraph should build AM14's form check instead.\nstrength: strong\n\n### F2\ntarget: docs/research/philosophy-of-the-unwritten-step.md:135\ntarget_words: \"required parts nothing is counted twice, and an established step moves the conclusion by one\"\nolder: docs/research/session26-open-rulings.md:896\nolder_words: \"a ground is no part, and grounds are read together, so they do not add\"\nkind: later-correction\nwhy: The 8 October ⚠ here says the grounds of a link's relevance claim are drawn as required parts. Row AD22 was RULED 2026-10-25 (\"All ruled, please build them\") and says that refinement is superseded. A ground is a SUPPORTS link carrying a step role, not a part, and grounds are read together by the weakest (Q38). His words: \"Should probably not be stored as parts\". The header's \"unruled\" for AD22 to AD29 is also out of date.\nstrength: strong\n\n### F3\ntarget: docs/research/writing-out-the-unwritten.md:580\ntarget_words: \"(in both orders; a disagreement goes to a person), by writing twice, and by looking for each sentence in the texts.\"\nolder: docs/research/session26-open-rulings.md:888\nolder_words: \"I don't think I ever blessed that particular part.\"\nkind: narrows\nwhy: AD14 is still OPEN. On 2026-10-10 the founder answered the \"writing twice\" check with \"'Asked twice'? I don't think I ever blessed that particular part.\" The row adds that it runs only in trial scripts, and that a resident note in CLAUDE.md that states it as a rule was a session's own. A reader ruling on AD14 should treat \"writing twice\" as unblessed, not as one of four agreed checks. AM14 later ruled against asking twice.\nstrength: strong\n\n### F4\ntarget: docs/research/what-counts-as-evidence.md:132\ntarget_words: \"standing, attacks on it, and whether it duplicates another answer do not matter.\"\nolder: docs/research/weighing-answers-and-the-badge-threshold.md:41\nolder_words: \"by its latest answer, whichever pole it took. An answer with no\"\nkind: later-correction\nwhy: §1b describes counting: +0.15 per answer, with an answer's standing, attacks and duplication ignored. Q3 was built on 2026-10-24. An answer now moves its pole by its own standing times relevance, one account counts once per question, and an attacked answer drops out. Ten answers from one account count as one. One first yes moves a claim to 0.501, where counting gave 0.506. §1b and the \"only Q3 ruled\" header describe code that no longer exists.\nstrength: strong\n\n### F5\ntarget: docs/ux-principles.md:97\ntarget_words: \"All contributions shift QBAF badge strength via QEM gradual semantics (Potyka, KR 2018).\"\nolder: docs/research/weighing-answers-and-the-badge-threshold.md:19\nolder_words: \"reason, any *no*, and a *yes* about the link itself all moved nothing; a *yes* to a question about the reason's\"\nkind: later-correction\nwhy: The public page (P4 and the \"works exactly as described\" caveat) says answering a critical question moves strength and ends in a changed badge colour. Measured 2026-10-24: a vote, a new reason, any no, and a yes about the link all moved nothing. A yes about the premise moved a claim to 0.506. Since then a badge waits for three weighed contributions (BADGE_MIN_WEIGHED). One answer therefore never changes a badge colour, so the sentence is stronger than the record supports.\nstrength: strong\n\n### F6\ntarget: docs/research/synthesis-build-plan.md:238\ntarget_words: \"whether, after reading some claim readings one at a time, a rule he blesses lets the rest be published in bulk.\"\nolder: docs/research/three-minds-and-the-leap.md:313\nolder_words: \"regularities that can help sift through and winnow out what remains to be ruled about the other cases\"\nkind: already-answers\nwhy: On 2026-10-19, the day before AP2, the founder said how he reads. He takes one or two cases at a time, in depth, and the point is to find regularities so that he can \"merely bless the application of some kind(s) of automated judgment of (ideally) all cases\". The canon line was blessed (Y9). That is his stated ideal for exactly the question AP2 leaves open. The cards should be chosen on purpose, and each reason he gives should be tried as a candidate rule on the rest.\nstrength: strong\n\n### F7\ntarget: docs/research/writing-out-the-unwritten.md:539\ntarget_words: \"A person reads the thirty written-out steps before any rate is claimed\"\nolder: docs/research/three-minds-and-the-leap.md:297\nolder_words: \"do it this way instead of using the sitting/review-pages, however amazing it was that they were\"\nkind: narrows\nwhy: §9.8 and AD20 propose a person reading all thirty steps with four fixed answers, and a sitting page was built for that. On 2026-10-19 the founder said he prefers to deep-dive one case at a time, writing his take before the next, \"instead of using the sitting/review-pages\". He wants to read one or two cases and then bless a rule. Someone ruling on AD20 should know the intended reading is a conversation per case, with the sitting page kept as an instrument.\nstrength: strong\n\n### F8\ntarget: docs/research/what-confirmed-means.md:116\ntarget_words: \"follows the same path, into the trust ladder's state or out.\"\nolder: docs/research/session26-open-rulings.md:1123\nolder_words: \"is retired as `confirmed` was: written on every link by its writer (high for extraction and people,\"\nkind: already-answers\nwhy: The note leaves the trust tier's fate open: \"into the trust ladder's state or out\". Row AQ2 settles it. The founder ruled \"Retire trust_tier, sure\" on 2026-10-20. It was written on every link, read by nothing that decides, and held nothing provenance does not. It is built and dropped from the live graph, and a guard refuses the name. The open end of §3 and §6 is closed, with no ladder state involved.\nstrength: strong\n\n### F9\ntarget: docs/research/philosophy-of-the-unwritten-step.md:506\ntarget_words: \"the rule can turn a working agreement into a dispute. The corpus has the general form in\"\nolder: docs/research/moral-convergence-metaethics-and-psychology.md:107\nolder_words: \"Maritain famously reported that the drafters could agree on the rights\"\nkind: same-idea-earlier\nwhy: §11 warns that demanding the rule can break an agreement that works, and cites Toulmin's commission and Sunstein. The older review already holds a famous instance. The drafters of the Universal Declaration agreed on the rights \"provided no one asks us why\". It is a third, independent case of agreement at the step with disagreement on the principle. It is also an instance of the mid-level convergence without foundational convergence that the corpus's convergence wager rests on.\nstrength: medium\n\n### F10\ntarget: docs/research/cognitive-bias-codex-and-human-contribution.md:141\ntarget_words: \"proposal to weight contributions by a contributor's demonstrated rationality is measuring noise\"\nolder: docs/research/what-counts-as-evidence.md:675\nolder_words: \"out-of-sample in 26 of 33 studies (Colson & Cooke 2017); in the IDEA protocol, status and experience\"\nkind: narrows\nwhy: The \"empirical kill\" on scoring people rests on bias-task reliability (0 to .82). The evidence note reports a different measure that works. Experts weighted by calibration on checkable questions beat equal weighting in 26 of 33 studies. It is limited to the fact layer, where ground truth arrives. So the kill covers bias-proneness and rationality scores, not calibration on resolvable claims. Anyone writing \"no contributor scoring, ever\" should keep that carve-out.\nstrength: medium\n\n### F11\ntarget: docs/ux-principles.md:468\ntarget_words: \"Use four independent signals (calibration, argument quality, intellectual honesty, evidence contribution) — never aggregated.\"\nolder: docs/research/what-counts-as-evidence.md:677\nolder_words: \"improving accuracy (Lorenz et al. 2011). Calibration can only be measured where ground truth arrives —\"\nkind: narrows\nwhy: The anti-pattern recommends a calibration signal and three other per-person signals as the alternative to one reputation number. The evidence note says calibration is measurable only where ground truth arrives. It is \"a route for the fact layer, never for evaluative claims\", and seeing others' estimates does not improve accuracy. So the calibration signal applies to resolvable claims only. The rest of the four-signal list is also unmeasured. The anti-pattern's alternative is stronger than the record supports.\nstrength: medium\n\n### F12\ntarget: docs/research/what-confirmed-means.md:50\ntarget_words: \"A sixth state, unset, was chosen on 2026-09-21 for machine splits, so that a part the splitter wrote\"\nolder: docs/ux-principles.md:378\nolder_words: \"A sentinel value meaning *nobody has looked* must never be routed into a comparison that assumes *a measurement*.\"\nkind: same-mechanism-elsewhere\nwhy: The same mechanism in a display. P22 shows a one-dimensional scale with no abstain state forcing nothing-answered onto the positive side: 115 of 116 claims, for five months. Here the true/false field had no way to say \"nobody looked\", so unset had to be invented. Both are cases of a representation that cannot abstain. P22's fix, a designed state for no reading, is what the `confirmed` retirement did with trust_state and human_action. The note could cite P22 as precedent.\nstrength: medium\n\n### F13\ntarget: docs/research/the-kind-of-mistake-we-make.md:42\ntarget_words: \"The mistakes that lived longest and cost most are all of the first shape, and each was invisible in\"\nolder: docs/ux-principles.md:380\nolder_words: \"115 of the 116 claims carrying any structure displayed confidence with nothing answered, for five months.\"\nkind: supplies-a-missing-piece\nwhy: §3 lists the long-lived, costly muddles. It omits a clean instance of the first shape. Four badge states (none addressed, some, all, failed) were compressed into three numeric bands. Nothing answered then computed to 0.5 and landed on the positive side, and no test could see a state with no representation. It ran for five months and failed silently. The page calls it a failure of vocabulary for \"no reading\". It also supports §9c guard 4 (define a property where it is written).\nstrength: medium\n\n### F14\ntarget: docs/research/synthesis-build-plan.md:113\ntarget_words: \"an LLM is never invited to grade its own omissions\"\nolder: docs/research/the-scrutiny-gap.md:64\nolder_words: \"the single most self-serving number in the system — the pipeline grading its own output\"\nkind: same-mechanism-elsewhere\nwhy: The same decision was taken earlier for extraction. The scrutiny-gap note found that the pipeline grading its own output was the most self-serving number, with no rubric and no anchors. In §3a it removed overall_quality_score and derived the counts in code. The synthesis build reaches the same split for omissions. The model judges which claims the answer rests on, and code does the arithmetic. This is a second instance of one rule: counts and grades come from code, not the model.\nstrength: medium\n\n### F15\ntarget: docs/research/synthesis-build-plan.md:180\ntarget_words: \"an omissions ledger is only complete up to the earliest truncation it can see\"\nolder: docs/research/world-completeness-and-its-relatives.md:16\nolder_words: \"What *can* be recorded is the search: what was looked for, by whom, with which moves, and\"\nkind: same-mechanism-elsewhere\nwhy: The omissions ledger reports what a budget dropped. It is not a statement about the world. It misses any stage above it, such as the 20-seed cap. The world-completeness note states the general law: absence of a record is not a record of absence. Only the search, including when it stopped, can be recorded. This says what the ledger may claim (\"these were in scope and not used\"). It must not say \"these are all the omissions\", and every cut above the prompt must be listed.\nstrength: medium\n\n### F16\ntarget: docs/ux-principles.md:306\ntarget_words: \"prioritize content that maximally updates THIS user's beliefs, not content with highest raw engagement\"\nolder: docs/research/the-feed-reconsidered.md:28\nolder_words: \"It is one list of claims, the same for every visitor, in eight\"\nkind: narrows\nwhy: P17 states the feed's properties: a quality threshold, surprise ranking for this user, and rankings that explain why. The feed measured on 2026-10-22 is one list, identical for every visitor, with eight sort orders. Its bridging mode returned one item and its weakest mode scored every claim 0.5. Nothing is personalised. A reader of the public page would assume personalised ranking exists, when it is a design intention only.\nstrength: medium\n\n### F17\ntarget: docs/research/philosophy-of-the-unwritten-step.md:297\ntarget_words: \"the principled reading of *what recurs is offered*. It also means that an assumption shared by\"\nolder: docs/research/saturation-and-the-long-tail.md:65\nolder_words: \"tail item by construction) and the DeepMind steering result into one picture: **frequency and value\"\nkind: same-mechanism-elsewhere\nwhy: The saturation note says frequency and value are anti-correlated in the region where machine compression drops items. It treats the implicit crux of a debate as a tail item. Offering a rule because it recurs across links (AF1, \"a group of links prompts the rule\") will surface the commonplace assumption first. The rare, load-bearing unsaid will come last or never. This is the same flattening mechanism in a different subject. It argues for a separate route to rare rules, such as a reserved place like the one reserved for rare claims in readings.\nstrength: medium\n\nNoticed: Ledger rows are the strongest source of rulings, but what-counts-as-evidence.md and ux-principles.md both name session26-open-rulings.md, so I could not use ledger rows against them. Several of the ledger's own rulings on Q3, Q18 and Q8 postdate what-counts-as-evidence. What-counts-as-evidence.md:3-7 says only Q3 is ruled \"in direction\" (it is now built). The Q18 \"open\" status (§8j) was pushed back on by the founder (2026-09-27, \"Questions are a first hand entity in the ontology of Deliberus, no?\") and the row proposes withdrawing the status. Q8 records his wish to keep the methodology ladder as a labelled starting point. CLAUDE.md:468 already makes the-kind-of-mistake §11 proposal 2 (small exact models) a resident rule, so §11's \"unruled\" is stale, but CLAUDE.md is not a corpus note. I ran one stray command that created an empty /tmp/dummy by accident, and I removed it at once. I wrote no other files.\n"}