{"path":"research/corpus-graph/readers/2026-10-27-grown-2-passes-proofs-and-paying-as-a-reader-returned-it.md","content":"# What a reader found to bear on eight grown notes on machine passes, their proof and who pays (2026-10-27)\n\nA read-only reader (Sonnet), sent through notes grown since their last reading (README § 13). It was given eight notes and\nread one in full (`docs/research/auto-connect-upgrade.md`); the other seven it read in their opening, newest dated sections\nand ⚠ lines only, so only the first is recorded as read and the seven stay owed:\n\n- `docs/research/pass-quality-and-the-choice-of-model.md` (not read whole)\n- `docs/research/proving-a-pass-reliable.md` (not read whole)\n- `docs/research/whose-tokens-pay-for-mapping.md` (not read whole)\n- `docs/research/machine-work-as-a-standing-property.md` (not read whole)\n- `docs/research/jev-and-the-judgement-layer.md` (not read whole)\n- `docs/research/microtasks-in-broad-daylight.md` (not read whole)\n- `docs/research/quality-control-for-deeper-steps.md` (not read whole)\n\nIt made 80 tool calls. Its report is below, unchanged. What was kept from it is in `../found.jsonl` (lines whose\n`where` names this file).\n\n---\n\n1 of the 8 notes read top to bottom (auto-connect-upgrade.md). The brief says \"seven\" but lists eight. The other seven I read in their opening answer, their newest dated sections and ⚠ lines, and the sections their findings rest on, not every line.\n\n### F1\ntarget: docs/research/auto-connect-upgrade.md:249\ntarget_words: \"more verdicts that will have to be read again. **This is the one that gates a run.**\"\nolder: docs/research/session26-open-rulings.md:574\nolder_words: \"his 2026-09-30 hold (*\\\"I'm still absorbing everything on both issues, don't feel ready for a decision yet.\\\"*) is lifted for this row\"\nkind: later-correction\nwhy: The note lists T5 (one definition per key) as open on the founder's own \"not ready\" and as the item that gates a keeper run. The ledger row for T5 is RULED 2026-10-24 \"for now\". His hold is lifted for this row, and the four-yes/no derivation (Z4, W7) stays open. A reader would stop treating T5 as the blocker and would look at item 2 and the other items instead.\nstrength: strong\n\n### F2\ntarget: docs/research/auto-connect-upgrade.md:101\ntarget_words: \"The menu has three relations (support, attack, qualification) and says that a premise the other claim depends on is support, with the premise as source.\"\nolder: docs/research/session26-open-rulings.md:1335\nolder_words: \"comes off the menus of the linking step, the keeper's relational half, extraction's relationship pass and the backwards-parts re-reading in one commit\"\nkind: later-correction\nwhy: Ledger BB2 is RULED. *Qualifies* comes off the menus of the linking step and of the backwards-parts re-reading in one commit, and each case it held goes elsewhere (support, attack, sameness key, or an attack on the relevance claim). The note's repaired menu, and its table of 25 pairs \"read again as qualification\", describe a menu that no longer stands. Anyone rebuilding or re-reading those links should use BB2's mapping.\nstrength: strong\n\n### F3\ntarget: docs/research/microtasks-in-broad-daylight.md:609\ntarget_words: \"**Proposed, unruled (ledger T9).** The splitter asks one more question of a cut, before the parts:\"\nolder: docs/research/session26-open-rulings.md:578\nolder_words: \"**RULED 2026-10-17**, founder, after the floor step had held every claim with an inline reason until this existed: *\\\"Build T9 then pls\\\"*\"\nkind: later-correction\nwhy: The note presents the reasons-cut as a proposal and says \"Nothing was changed\". The ledger row T9 is RULED 2026-10-17 and was built the same night (`reasons_off.py`). The row also sharpens the headline: the cut keeps its pieces and adds the support links, instead of \"never as parts\". A reader about to build or review this should treat it as live and use the sharpened form.\nstrength: strong\n\n### F4\ntarget: docs/research/microtasks-in-broad-daylight.md:37\ntarget_words: \"part caps its parent, refuting any of the six would pull the parent down, which is backwards.\"\nolder: docs/research/session26-open-rulings.md:1388\nolder_words: \"A part link records a cut and carries no strength: nothing a part says reaches its whole, neither the weakest-part cap\"\nkind: narrows\nwhy: The splitter's harm argument rests on a written part capping its parent. Ledger BD16 (RULED 2026-10-25) says part links carry no strength: the cap is switched off in the code and the arithmetic is kept only behind a flag. The six harmful parts still speak for the author in his name. They no longer pull a parent's strength down. Anyone ranking the splitter's harms or setting its error tolerance should drop the strength-cap rationale.\nstrength: strong\n\n### F5\ntarget: docs/research/pass-quality-and-the-choice-of-model.md:66\ntarget_words: \"an hour of his in a sitting buys more than any change of model\"\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: The note makes a sitting the person's one-time labelling step, and the answer key its lasting asset (§ 5e step 4). On 19 October he said he does not want to read in sittings. He wants one case at a time, in depth, cases chosen on purpose, and the canon line was re-cut and blessed. The labelled-sample step and Y2/Y3 stay, but the way the labels are produced changes, and the draw-selection bias he flags matters. This is not in the note.\nstrength: strong\n\n### F6\ntarget: docs/research/pass-quality-and-the-choice-of-model.md:332\ntarget_words: \"Of the 25 places in the code that ask a model, three keep a record of which model answered\"\nolder: docs/research/storing-the-record-of-machine-judgments.md:14\nolder_words: \"one row for every call to a model, saying which model answered, under which\"\nkind: later-correction\nwhy: The note says that for 22 of 25 places \"which model wrote this\" has no answer. The next day (2026-10-06) the founder ruled \"Yes build it\" on the record of machine judgments, and it went live that evening. Every call is now one write-once row naming the model that answered. The measured gap, and proposal item 1 (\"every pass that writes stores which model answered\"), are largely overtaken. Only the pointer from each written item to its row remained open (Z1b).\nstrength: strong\n\n### F7\ntarget: docs/research/quality-control-for-deeper-steps.md:120\ntarget_words: \"often an answer is wrong. No small task here has an error rate measured against a person, and the\"\nolder: docs/research/auto-connect-upgrade.md:259\nolder_words: \"confidence* reads as measured calibration. He has read nine keeper pairs (2026-10-18, through\"\nkind: later-correction\nwhy: The 10 October count (zero answers, zero notes, control 6 \"nothing\") is dated. By 18 to 19 October he had read nine keeper pairs, and one in depth (a three-minds pair). The record has no field for his take, so this is a reading, not a stored label. The \"nothing\" row and the claim that the proofs and ladder have nothing to stand on should be softened. Also note he reads one case at a time, not in sittings.\nstrength: strong\n\n### F8\ntarget: docs/research/auto-connect-upgrade.md:91\ntarget_words: \"still had to put a word in it, and the word models put there was the first on the menu\"\nolder: docs/research/session26-open-rulings.md:445\nolder_words: \"such as a default, a value borrowed from a neighbouring event, an example value, or a relation read off another label\"\nkind: same-idea-earlier\nwhy: The 47 false merge marks are a plausible stand-in written where a blank belonged: a required field the model had to fill, filled with the first menu word. Ledger Q30 (RULED 2026-09-28, added globally) already states the guardrail \"store the blank, never a plausible stand-in\" and lists five earlier instances. The note records this as a new find. Q30 would have predicted it, and it is a sixth instance to log. The same fix applies to any other required field behind a gate.\nstrength: strong\n\n### F9\ntarget: docs/research/auto-connect-upgrade.md:274\ntarget_words: \"The keeper reads two texts and nothing else: not the kinds, not a `REPORTS` link between\"\nolder: docs/research/session26-open-rulings.md:1046\nolder_words: \"The keeper is handed each claim's quoted passage and the senses in play for a shared term\"\nkind: later-correction\nwhy: The note's item 7 (written 18 October) says the keeper sees only two texts. Ledger AM9 was RULED 2026-10-19: it is handed each claim's quoted passage plus the senses in play for a shared term, with the passages marked off in tags and the anchoring worry to be measured on the nine stickprov pairs. That narrows item 7. It does not give the keeper the kinds or the REPORTS link, so the rendering-pair question stays open. It also changes what \"a keeper verdict\" means before and after.\nstrength: strong\n\n### F10\ntarget: docs/research/microtasks-in-broad-daylight.md:632\ntarget_words: \"splitter that files reasons right repairs the 93; extraction that never bundles them is the\"\nolder: docs/research/session26-open-rulings.md:1021\nolder_words: \"All claims need to be split using the splitter when entering Deliberus. Then they need to be sameness-clustered/merged, after splitting, not before\"\nkind: already-answers\nwhy: The note says preventing bundling at extraction \"is not proposed here\". Ledger AL9 (RULED 2026-10-17) rules exactly that order: every claim is cut to the separability floor first, then sameness, then links. It was built behind DELIBERUS_FLOOR_FIRST. AX8 (23 October) rules every claim is stored at the floor. The upstream prevention is therefore ruled, and the splitter's repair of the 93 is a retrofit, not the only fix.\nstrength: strong\n\n### F11\ntarget: docs/research/microtasks-in-broad-daylight.md:655\ntarget_words: \"1. **One definition per key, in all three places.** Rewrite the sitting's key texts from the\"\nolder: docs/research/session26-open-rulings.md:574\nolder_words: \"his 2026-09-30 hold (*\\\"I'm still absorbing everything on both issues, don't feel ready for a decision yet.\\\"*) is lifted for this row\"\nkind: already-answers\nwhy: The section heading says \"(unruled)\". Item 1 is ledger T5, now RULED 2026-10-24 \"for now\": one definition per key, identical in the instructions, on the review page and in the routing code. Items 2 to 6 are not covered. Item 2, the four yes/no derivation, is still open as Z4 and W7. A reader should build item 1 as ruled and treat the rest as unruled.\nstrength: strong\n\n### F12\ntarget: docs/research/pass-quality-and-the-choice-of-model.md:46\ntarget_words: \"4. **The model's tier is the third lever.** Of nineteen measured failures of machine passes in\"\nolder: docs/research/three-minds-and-the-leap.md:160\nolder_words: \"clear and still be answered badly when the answer needs an object the reader does not have.\"\nkind: narrows\nwhy: The note says most measured failures trace to how the question was posed, so a better question fixes them. three-minds takes the keeper's overlaps-by-topic answer, which has the clause \"judged by EVIDENCE, not by topic\" in capitals in its prompt. It states the limit of that finding: an instruction can be clear and still answered badly when the answer needs an object the reader lacks. A reader should not conclude that tightening wording is enough. The reader may need the passage, the sense, the chain.\nstrength: strong\n\n### F13\ntarget: docs/research/microtasks-in-broad-daylight.md:28\ntarget_words: \"2. **Our tasks are better specified, and the variance that remains is mostly ours to remove.** It\"\nolder: docs/research/three-minds-and-the-leap.md:160\nolder_words: \"clear and still be answered badly when the answer needs an object the reader does not have.\"\nkind: narrows\nwhy: The note reads the keeper as \"well specified\" with the evidence-not-topic clause in the prompt. It locates the remaining variance in the three-places definition split. A measured case after the note: the keeper had that clause and still answered overlaps by topic, so the instruction was clear and the answer was bad. Fixing the definitions (now T5) is necessary but not sufficient. The reader also needs the object the verdict depends on, which is where AM1/AM9 go.\nstrength: medium\n\n### F14\ntarget: docs/research/jev-and-the-judgement-layer.md:362\ntarget_words: \"1. Do the two bear on each other at all? No → *unrelated*.\"\nolder: docs/research/what-a-record-linkage-project-teaches-deliberus.md:255\nolder_words: \"The last row should read *no sameness relation*, because relatedness should be asked once.\"\nkind: supplies-a-missing-piece\nwhy: § 3g derives the key from four yes/no questions and puts \"do they bear at all\" inside the keeper, with \"otherwise overlaps\". The record-linkage note (§ 4c, § 4d) takes the same table and changes two things. The fourth question must be \"is there a statement both assert\", or overlaps absorbs the roughly 64 reason-for pairs. Relatedness must be asked once, by the relation classifier, not twice. Anyone trialling the four questions should start from that corrected table, not § 3g's.\nstrength: medium\n\n### F15\ntarget: docs/research/pass-quality-and-the-choice-of-model.md:663\ntarget_words: \"4. **Y4. Writer and checkers from different model families**, and a write needs every checker's\"\nolder: docs/research/the-scrutiny-gap.md:102\nolder_words: \"the two signals must differ in mechanism or in model family\"\nkind: same-idea-earlier\nwhy: The August scrutiny-gap note already states Y4 as a condition on co-stimulation: two instances of one model are one detector sampled twice, with the measured 18 of 30 agents naming the same branch. It also names the stance instrument (cosine against lexical syntax) as a working example of two different mechanisms, which the pass-quality note lacks. That strengthens the case for different kinds of checker, not only different makers, and shows Y4 extends an existing rule.\nstrength: medium\n\n### F16\ntarget: docs/research/whose-tokens-pay-for-mapping.md:683\ntarget_words: \"- **A rate per person**, above the per-minute limits that exist.\"\nolder: docs/research/the-scrutiny-gap.md:106\nolder_words: \"A per-daemon budget does not catch this; a global rate-and-spend ceiling does.\"\nkind: supplies-a-missing-piece\nwhy: The note replaces the removed spend gate with per-person limits, a per-debate count and drafts. The scrutiny-gap note measured a runaway population (2.4 million job requests, 117 accepted) and concluded that per-actor budgets miss it and a global rate-and-spend ceiling is needed. With many contributors on their own keys, volume is exactly such a population. A global ceiling should sit beside X3.\nstrength: medium\n\n### F17\ntarget: docs/research/machine-work-as-a-standing-property.md:608\ntarget_words: \"2. *A standing pass costs no more calls than its door.* The condition *never read* holds once per item,\"\nolder: docs/research/auto-connect-upgrade.md:281\nolder_words: \"pairs outside the five candidates above): a run over \\\"every pair\\\" is a different step from the\"\nkind: narrows\nwhy: Principle 2 assumes a standing condition owes what its door reads. For the keeper the door reads each new claim's five nearest candidates, while the standing condition is every resemblance link. After the machine-consciousness run, 11 of 192 pairs at 0.80 or more were outside every run's candidates. The note says a run over \"every pair\" is a different step with its own ceiling. A builder should expect the owed count to exceed the door's cost and plan the cap accordingly.\nstrength: medium\n\n### F18\ntarget: docs/research/proving-a-pass-reliable.md:359\ntarget_words: \"verdict three ways and read their agreement as the confidence, at three times the model calls.\"\nolder: docs/research/jev-and-the-judgement-layer.md:64\nolder_words: \"A pre-registered evaluation found accuracy **flat from 0.50 to 0.95 confidence**, then jumping to\"\nkind: supplies-a-missing-piece\nwhy: § 5 builds the abstention guarantee on a confidence the pass can produce, and suggests agreement across three wordings. The Jev note reports calibration results for a purpose-built decision model. Out of distribution its error is 4.4 times the noise floor, accuracy is flat across 0.50 to 0.95 stated confidence, and on unknowable tasks it was right 44.7% at 0.74 stated. Any threshold must therefore be calibrated per question against a person's labels, not read from stated confidence. The abstention route yields no bound until then.\nstrength: medium\n\nNoticed: Another session was editing pass-quality-and-the-choice-of-model.md while I read (it grew by 2 lines, and line numbers in F5 and F6 are as of my last check); re-verify the quotes before use. I did not report ledger-vs-ledger or ledger-to-target pairs for pass-quality, proving or machine-work, because those three name session26-open-rulings.md. Several other ledger rows touch the same ground: W4 stays OPEN and AB12 stays \"open as to how\". Rows on the keeper's open points (T5, AM9) were ruled after the 18 October \"what is open\" list in auto-connect-upgrade.md and that list is now stale in items 1 and 7.\n"}