{"path":"research/corpus-graph/readers/2026-10-27-grown-2c-judgement-layer-microtasks-and-deeper-checks-as-a-reader-returned-it.md","content":"# What a reader found to bear on three grown notes on the decision model, microtasks in broad daylight and quality control for deeper steps (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 3 notes:\n\n- `docs/research/jev-and-the-judgement-layer.md`\n- `docs/research/microtasks-in-broad-daylight.md`\n- `docs/research/quality-control-for-deeper-steps.md`\n\nIt made 98 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 3 of the 3 notes the brief lists in full (the brief says \"seven\" but names only three), every line top to bottom. None of the three is a public overview page under docs/, so there are no public-page findings. All 15 findings below were opened at both ends, and a script confirmed each quote is on a single line of its file. I checked found.jsonl before choosing pairs and left out any pair already kept there.\n\n### F1\ntarget: docs/research/jev-and-the-judgement-layer.md:678\ntarget_words: \"person adds presumes relevance 0.5 under the build. A confident Jev *does not bear* could lower that\"\nolder: docs/research/checks-on-machines-and-people.md:522\nolder_words: \"For a link the reading is a good question and a poor switch\"\nkind: narrows\nwhy: The Jev note floats letting a confident \"does not bear\" drop a person-added link's presumed relevance to 0. The later note replayed the sameness reader over the 32 reasons and objections people actually filed. It would withhold 2, both honest legal objections attached to the wrong claim, because the page offers no place to object to a step. Its verdict for links is to show the reading as a question to the author, not to use it as a switch. That bears directly on ledger W5.\nstrength: strong\n\n### F2\ntarget: docs/research/jev-and-the-judgement-layer.md:486\ntarget_words: \"Sacredness brake | lexicon matched **0 of 1,678** claims | yes/no: *does this claim declare a value not open to trade-off?*\"\nolder: docs/research/the-load-bearing-unsaid.md:120\nolder_words: \"Measured effect: 0 → 2 firings on the corpus, both genuine author value premises\"\nkind: later-correction\nwhy: The Jev table gives the lexicon's \"today\" as matching nothing. That was the August state before the lexicon was widened. On 13 August the brake was rebuilt, with a sanctity register and reported-speech syntax suppression, and it now fires on 2 claims, both genuine author value premises. The note also records that an attributed_to guard was killed by measurement. Someone ruling on the \"Park\" verdict, or on whether a model detector is needed, would be weighing a stale baseline.\nstrength: strong\n\n### F3\ntarget: docs/research/jev-and-the-judgement-layer.md:501\ntarget_words: \"ruled 2026-09-13, re-check unbuilt | yes/no per edge: *does A actually support B?*\"\nolder: docs/research/session26-open-rulings.md:527\nolder_words: \"a confident later *unrelated* from the keeper (no overlap refused on its chain) or a confident *same* retires the older machine links\"\nkind: later-correction\nwhy: Jev calls the ruled confident re-check \"unbuilt\" and proposes Jev's raw extremes (0.9 and 0.1) as its trigger. Ledger R12, ruled 2026-10-19, says the re-check was built that day on the keeper's confident readings. It also says \"high confidence\" in the housekeeping ruling means measured calibration, not a pass's own confidence, and the founder's reason was reversibility. Jev's note itself finds accuracy flat from 0.50 to 0.95, so its extremes would not meet R12's reading.\nstrength: strong\n\n### F4\ntarget: docs/research/quality-control-for-deeper-steps.md:466\ntarget_words: \"A sitting for steps: one step beside its\"\nolder: docs/research/session26-open-rulings.md:728\nolder_words: \"So I think I wanna do it this way instead of using the sitting/review-pages, however amazing it was that they were created for me.\"\nkind: narrows\nwhy: The note plans a sitting for steps, drawn at random across levels with one item in five blind, and builds the page for it. Ledger Y9 records his blessed gist of 2026-10-19: he reads one case at a time and writes his take before the next. He prefers that to the sitting and review pages, and he worries about bias in which cases come first. A person about to rule on AG11's order, or on the page's design, should read this first.\nstrength: strong\n\n### F5\ntarget: docs/research/quality-control-for-deeper-steps.md:369\ntarget_words: \"writes the chain, one stage to a sentence, as questions; looks every stage up, on both sides\"\nolder: docs/research/session26-open-rulings.md:953\nolder_words: \"a step is at its floor when it applies *one stated rule, once, to one stated case*, so the two tests reach the smallest step\"\nkind: already-answers\nwhy: The note writes a chain of \"stages\" for how-it-comes-about answers, with 4 to 5 stages per chain. It has no stated size for a stage, and it admits that three of eight chains were arguments or arithmetic set out in stages. AG14 was ruled on 2026-10-13 and 2026-10-14 (founder: \"Excellent. Yes, agreed.\" and \"Yes that's good too, that's what I want.\"). It sets the standard for anything a machine writes under a link: smallest step, no gap, nothing removable, one rule applied once to one case. That is the unit these chains should be cut to.\nstrength: strong\n\n### F6\ntarget: docs/research/microtasks-in-broad-daylight.md:588\ntarget_words: \"A part and a reason are stored on two kinds of link,\"\nolder: docs/research/checks-on-machines-and-people.md:558\nolder_words: \"own reasons were filed as its pieces, here a person's reasons are filed as its parts.\"\nkind: same-mechanism-elsewhere\nwhy: § 3h finds the splitter filing a sentence's own reasons as parts, where a part caps its claim. The later note finds the same fault at a second door. People answering the page's \"What's underneath?\" box give what the claim rests on, and it is stored as what the claim consists of. 28 of 39 person-attached parts read as not changing the claim. So the fault is not only the splitter's. Fixing T9 in the splitter leaves the box's door open, and a ruling on T9 should name both.\nstrength: strong\n\n### F7\ntarget: docs/research/jev-and-the-judgement-layer.md:517\ntarget_words: \"It fills a gap the threat model names and no instrument covers, it is a telemetry reading rather than content on a claim\"\nolder: docs/research/a-sentence-as-a-pointer-into-a-landscape.md:355\nolder_words: \"one request a claim, one yes/no question a word, *does this word say\"\nkind: already-answers\nwhy: Jev proposes amplification, a claim carrying more weight than its source, as \"the most novel placement\", still waiting for labels. The pointer note records an instrument of that family, already built and tried. It is added_words.py, a one-request-per-claim decision-model check on words the rewrite added. It was founder-approved on 2026-10-19, scored the hurdle claim's \"significant\" at 0.87 to 0.92, and sits behind the same one-case-first review. The measured instance is \"a second issue\" becoming \"a significant technical hurdle\".\nstrength: medium\n\n### F8\ntarget: docs/research/quality-control-for-deeper-steps.md:391\ntarget_words: \"No sentence says a sentence above it, or either claim of the link, over again\"\nolder: docs/research/three-minds-and-the-leap.md:356\nolder_words: \"a chain whose steps only restate its two ends passes.\"\nkind: supplies-a-missing-piece\nwhy: The note proposes a circle check, but only as a plan, and its trial never produced a circle. The keeper's chain for the J-space pair was a real case. Both steps were asserted and the step count passed, yet step one restated claim A and step two restated claim B, so nothing sat between them. The proposed remedy is the same as the note's: a step that is the same claim as either end does not count. This is a measured instance and a ready test case.\nstrength: medium\n\n### F9\ntarget: docs/research/quality-control-for-deeper-steps.md:611\ntarget_words: \"So this check cannot tell a complete chain from one with a hole. Nothing built can.\"\nolder: docs/research/the-chain-inside-a-link/ten-links-written-to-the-floor.md:42\nolder_words: \"holds and the end fails, it found a gap in 13 of the same 25 steps.\"\nkind: supplies-a-missing-piece\nwhy: The note concludes that no built check can find a hole in a chain. Three days later, a trial on ten links showed that the question put to the reader changes this. Asked \"is anything missing?\", the reader passed every step. Asked for a situation where every sentence holds and the end fails, it found a gap in 13 of 25 steps. That reader is not the first-level rule-picker the note tried, so \"nothing built can\" is stronger than the record now supports. The same trial warns that this reader flips on 6 of 25 steps when asked twice.\nstrength: medium\n\n### F10\ntarget: docs/research/jev-and-the-judgement-layer.md:484\ntarget_words: \"Most claims are not reported speech, so most writing calls disappear.\"\nolder: docs/research/earning-the-right-to-write.md:483\nolder_words: \"A 30% mint rate is the pre-filter working as designed, not a poor one.\"\nkind: narrows\nwhy: Jev proposes a yes/no gate in front of the reported-speech writing call so most writing calls vanish. The backfill note shows a cheap lexical pre-filter already does that gating. Of 1,859 claims only 153 were read, and 46 were minted. The model's refusals are already recorded so a repeat run skips them. A Jev gate would be a second gate behind the first. The saving is over roughly 153 reads, not 1,859, and 30% of the pre-filtered reads are real reported speech.\nstrength: medium\n\n### F11\ntarget: docs/research/jev-and-the-judgement-layer.md:485\ntarget_words: \"The 72 blind-labelled wild Swedish sentences are exactly the per-question, per-language set test 5 asks for.\"\nolder: docs/research/weighing-eval/ground-truth-public.md:15\nolder_words: \"SP12's W confirmed. Key totals now 6 W / 66 N.\"\nkind: narrows\nwhy: Jev counts the 72 labelled Swedish sentences as a ready set for fitting a per-question calibration. After adjudication the key holds 6 weighings and 66 non-weighings. Six positives cannot fit a probability-to-accuracy curve for a yes/no, and cannot show where the confident band is right. Jev's own § 6d notes that forty links bound a rate near 7%. The set suits a first recall check on the Swedish weighing question, not a threshold.\nstrength: medium\n\n### F12\ntarget: docs/research/jev-and-the-judgement-layer.md:252\ntarget_words: \"send both claims in full, never the labels.\"\nolder: docs/research/three-minds-and-the-leap.md:345\nolder_words: \"is a use-relative question too. Proposed as AM9: hand the keeper each claim's quoted passage and the\"\nkind: narrows\nwhy: Jev's rule from the outside tests is to send both claims in full. The keeper note shows that full claim text is not enough, because the claims are stranger-rewritten sentences that have lost their game. The nine-pair reading found the keeper's \"no bearing\" true of the sentences and false of the moves. The remedy (AM9) hands over each claim's quoted passage and the senses in play, and the founder said to go ahead, with the anchoring worry to be measured. So any Jev trial of relatedness should be given the passage and senses too.\nstrength: medium\n\n### F13\ntarget: docs/research/jev-and-the-judgement-layer.md:359\ntarget_words: \"claims can hold, so they split into yes/no questions, the shape in which Jev gained most (62.6%\"\nolder: docs/research/microtasks-in-broad-daylight.md:658\nolder_words: \"entail B? does B entail A? can both be true? do they share content? Each is closer to a checkable\"\nkind: same-idea-earlier\nwhy: § 3g presents the seven keys as four yes/no questions as a fresh proposal prompted by the founder, and calls it untested. The same decomposition, with nearly the same four questions, was proposed on 2026-09-29 as § 4 item 2, with the argument that each is closer to a checkable question. A later fixed table and two changes to it are in ledger Z4. That older version adds a reason beyond Jev's gains: it keeps the keeper's key checkable regardless of which model answers.\nstrength: medium\n\n### F14\ntarget: docs/research/microtasks-in-broad-daylight.md:262\ntarget_words: \"both runs, plus one part carrying added commentary; 45 of 45 parts were judged the same way twice.\"\nolder: docs/research/fragile-checkers-and-the-verification-bottleneck.md:35\nolder_words: \"The most mechanically stable model in the panel (2% flip rate under rephrasing) was the fourth most persuadable under sustained pressure (44%).\"\nkind: narrows\nwhy: The note leans on the part test answering the same way twice (45 of 45, § 4 item 6) as evidence that it is sturdy. The Jagged Judges work in the corpus shows that stability under re-asking is a different property from holding the verdict under pressure. The most stable judge flipped 44% of the time under pressure. So 45 of 45 shows that the repeat reading is steady. It does not show that the verdict is right or robust, and it is one family reading itself.\nstrength: medium\n\n### F15\ntarget: docs/research/quality-control-for-deeper-steps.md:194\ntarget_words: \"its context holds its own earlier errors, and that taking earlier output out of the context helps\"\nolder: docs/research/flywheels-in-building-deliberus/literature-report-2026-10-22.md:219\nolder_words: \"models are more likely to err when their own earlier mistakes sit in context\"\nkind: supplies-a-missing-piece\nwhy: The note flags its depth argument as unverified and cites a helper's reading of Sinha et al. A later literature report states the same finding. It also adds that larger models alone do not fix it and that thinking models mitigate it. That report read the paper at abstract or snippet level only, not the body, so it is a second statement of the same source and not independent confirmation. The \"thinking models mitigate\" clause matters for choosing the writer of a level.\nstrength: medium\n\nNoticed: Early on I mistakenly ran a shell command that created an empty file, /tmp/nonexistent_ok. I removed it at once, and the repository was not touched. The brief's \"1 of your notes are public overview pages\" does not apply, since none of the three is a public overview page under docs/. Many obvious pairs involving session26-open-rulings.md and microtasks-in-broad-daylight.md, and QC with the-chain-inside-a-link, are already in found.jsonl, so I left them out. Ledger row T5 was ruled on 2026-10-24 (\"for now\") and answers § 4 item 1 of the microtasks note, but that pair is already kept. Rows W1 to W7 are all still OPEN.\n"}