{"path":"research/corpus-graph/readers/2026-10-10-judges-models-and-alignment-as-a-reader-returned-it.md","content":"# What a reader found to bear on nine notes on machine judges, models and alignment (2026-10-10)\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 9 notes:\n\n- `docs/research/jev-and-the-judgement-layer.md`\n- `docs/research/whose-tokens-pay-for-mapping.md`\n- `docs/research/fragile-checkers-and-the-verification-bottleneck.md`\n- `docs/research/grounded-critics-and-quarantined-imagination.md`\n- `docs/research/world-models-and-the-ontology-revival.md`\n- `docs/research/steadying-an-outside-judge.md`\n- `docs/research/deliberus-as-alignment-infrastructure.md`\n- `docs/research/ai-safety-and-tao-augmentation-research.md`\n- `docs/research/sff-2026-and-fiscal-sponsorship-path.md`\n\nIt made 148 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 nine notes in full. Two things to know first. I created two small helper scripts in the session scratchpad (outside the repository) to check \"name occurs in target\" and \"pair in found.jsonl\". I touched no file in the repository. Other sessions edit these files, and one line number in `convergence-wager-typed-residues-synthesis.md` shifted by one while I worked, so recheck line numbers before use. All 36 quotes were matched programmatically to a single line at the moment of writing.\n\n### F1\ntarget: docs/research/grounded-critics-and-quarantined-imagination.md:39\ntarget_words: \"a machine judgment on contested ground lands as a proposal a human ratifies; wrong proposals are corrected at the boundary\"\nolder: docs/research/the-ratification-record-and-the-triage-feed.md:209\nolder_words: \"Machine discernments carry machine provenance forever, displayed as such,\"\nkind: later-correction\nwhy: On 2026-08-31, six days after this note, the founder blessed an end-state with no ratify act: assent is telemetry, dissent is a contribution, and verdict-acts do not exist. Machine discernments keep machine provenance forever, shown as such and open to challenge. So \"a human ratifies\", the earn-versus-certify framing in § 4 and the \"ratified transitions\" sentence proposed in § 7.3 describe a design he has since declined. Build the provenance-split readout on provenance, not on a ratification gate.\nstrength: strong\n\n### F2\ntarget: docs/research/fragile-checkers-and-the-verification-bottleneck.md:75\ntarget_words: \"we have never measured it: the same source re-extracted, or extracted under the capacity-fallback model swap, yields different claim sets\"\nolder: docs/research/cross-source-premise-run.md:92\nolder_words: \"The minimum-wage pair returned two premises on the first pass and three on the second, from identical inputs.\"\nkind: later-correction\nwhy: Two days later a pre-registered premise run produced the first instance. The note calls it extraction non-determinism measured on this pass, the mechanical-consistency tier the wiggle reading flagged as untested. It adds that a decomposition method that returns a different set on re-run has a stability number and nobody has taken it. So \"never measured\" is now one case, not zero. Whoever builds the § 6.3 paraphrase tier starts from a known variance of at least one premise in three.\nstrength: strong\n\n### F3\ntarget: docs/research/deliberus-as-alignment-infrastructure.md:101\ntarget_words: \"If human values converge when decomposed to sufficient depth, that convergence IS the alignment target.\"\nolder: docs/research/convergence-wager-typed-residues-synthesis.md:26\nolder_words: \"the wager is that the residue space is small, precisely stateable, still arguable in a different register, and empirically accounts for a minority\"\nkind: narrows\nwhy: In July the founder held convergence as a wager held fallibly, with legibility as the guaranteed floor. The synthesis moves the claim from \"values converge when decomposed far enough\" to \"decomposition leaves typed, narrow, legible residues\". It names the Forecasting Research Institute collaboration (80 median hours, mutual understanding reached, medians nearly unmoved) as the hard bound. A reader should pitch the residue map and its measured size as the alignment-relevant product, not convergence as the target.\nstrength: strong\n\n### F4\ntarget: docs/research/ai-safety-and-tao-augmentation-research.md:275\ntarget_words: \"The extraction pipeline already works. 318 claims, 93 cross-extraction connections, 372 CQs. The `@[simp]` flywheel is validated\"\nolder: docs/research/the-reuse-flywheel-prior.md:190\nolder_words: \"corpus size. It tracks whether an **adversarial counterpart is already present** — death-penalty pair\"\nkind: contradicts\nwhy: Measured 2026-08-29 on 31 sources, per-source cross-source connection yield does not rise with corpus size. It tracks whether an adversarial counterpart is already present, so \"the flywheel is paired, not cumulative\". Only 4 of 90 concepts span more than one source, and every connection came from the old cosine-only linker. \"Validated\" rests on 93 edges from seven March extractions. Do not carry that sentence into a funder-facing or public line.\nstrength: strong\n\n### F5\ntarget: docs/research/ai-safety-and-tao-augmentation-research.md:198\ntarget_words: \"A human-maintained argument graph is the antidote — genuine human reasoning, structured and persistent, immune to the recursive degradation that afflicts AI-generated content.\"\nolder: docs/research/major-transitions-heredity-and-evolvability.md:553\nolder_words: \"generation at a time, which is the shape of model collapse: tails thin, variance shrinks, and\"\nkind: contradicts\nwhy: The corpus's own code reading says a machine-cut part can be linked to and cut again with no person in between. Machine structure built on machine structure compounds same-hand bias a generation at a time, which is the shape of model collapse, and nothing throws an error. The proposal that followed was revised into ledger T1 (ruled 2026-09-29). The graph is therefore not immune by being structured and persistent. The antidote holds only where people supply the content.\nstrength: strong\n\n### F6\ntarget: docs/research/deliberus-as-alignment-infrastructure.md:25\ntarget_words: \"AI systems could verify their ethical reasoning against a Deliberus graph the way AlphaProof verifies mathematical reasoning against Lean\"\nolder: docs/research/lean-deliberus-analogies.md:184\nolder_words: \"Narrower: Lean's kernel pays **twice**, and only one of those payments is available here.\"\nkind: narrows\nwhy: The Lean note's 2026 sharpening says the kernel both checks and supplies free ground-truth reward, and deliberation has no such oracle, only validity of form. So borrow the checkability argument, do not expect the training flywheel, and do not model the trajectory on AlphaProof. A ⚠ line adds that a machine's yes or no on whether one claim is a reason for another changed with wording on 18 of 30 links, so relevance is no kernel check. \"This is not metaphor\" overstates.\nstrength: strong\n\n### F7\ntarget: docs/research/world-models-and-the-ontology-revival.md:35\ntarget_words: \"Run 6 found exactly such an accumulation at the relationship layer, which is why the stance instrument exists and why it reports candidates\"\nolder: docs/research/the-remaining-case-for-the-stance-guard.md:7\nolder_words: \"84 candidates split into 5 it was built for, 14 that are a *different* bug, and about 65 healthy shared ground.\"\nkind: later-correction\nwhy: Run 6's own 14 August correction says no such agreement edge was stored. On 2026-09-13 all 84 candidates were read: 5 were reported speech, 14 were false SUPPORTS edges of a different kind (mirrored vocabulary, different referents), and about 65 were healthy. The guard was demoted from a human-facing endpoint to a batch check, and a stance field plus an entailment re-check were ruled. The relationship-layer gap is real but far narrower than \"accumulation\", and it now has a ruled remedy.\nstrength: strong\n\n### F8\ntarget: docs/research/steadying-an-outside-judge.md:35\ntarget_words: \"twice — once with the model given a graph query tool, once without — and measure the flip rate under fabricated consensus.\"\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: In agent-memory research an agent that owns a store mostly does not consult it. A coding agent with a filled memory made no memory call in 114 turns, while the same items delivered by its harness at set moments arrived 138 of 138 times. The mechanism here, a judge that \"has a graph query tool\", may show no effect because the tool goes unused. The test design, and mechanism 1, need the lookup delivered at the moment of pressure, not merely offered.\nstrength: strong\n\n### F9\ntarget: docs/research/steadying-an-outside-judge.md:21\ntarget_words: \"balanced attack-and-support energy, a high hinge, an unclassified terminus, a `does_not_fit` scheme\"\nolder: docs/research/lowering-the-cost.md:184\nolder_words: \"Links with a standing question's claim at one end stand at 2,060 supports to 2,058 attacks, balanced by construction.\"\nkind: narrows\nwhy: Measured 2026-10-08, the near-balance of attacks and supports is scaffolding: question-claim links run 2,060 to 2,058 by construction. People have written 9 supports and 21 attacks, too few to read a tendency from. So the proposed fragility signal, \"accumulated human disagreement\", is today almost entirely machine-minted balance. It would flag scaffolded claims as contested and say nothing about the truly contested ones until human contribution exists.\nstrength: strong\n\n### F10\ntarget: docs/research/fragile-checkers-and-the-verification-bottleneck.md:143\ntarget_words: \"No adversarial-accessible surface may re-invoke a judge on a standing verdict with challenger content in its context\"\nolder: docs/research/jev-and-the-judgement-layer.md:684\nolder_words: \"other models and as lowering, as filed, by Claude both times. So four readers of two kinds agree\"\nkind: narrows\nwhy: The relevance judge, built a month after this note, is told how each link is filed, and that filing was written by a person or a machine. On the one link filed as an attack, Claude followed the filing twice. Three other models given the same prompt, and the decision model, which was not told, read it as support. A stored filing is standing-verdict content in the judge's context, so \"current code passes\" no longer holds there. The rule should say whether filings are covered.\nstrength: strong\n\n### F11\ntarget: docs/research/jev-and-the-judgement-layer.md:52\ntarget_words: \"that breaking a judgement into atomic sub-judgements improves the judgement — the same operation\"\nolder: docs/research/writing-out-the-unwritten/outside-reading-dependable-small-tasks.md:229\nolder_words: \"Weak verifiers gain. Strong ones can lose. More sub-claims help first, then hurt.\"\nkind: narrows\nwhy: An independent literature reading (Hu et al., NAACL 2025) reports the same asymmetry as the phishing result: decomposition lifts weak verifiers and can lower strong ones. F1 fell from 72.32 to 59.90 with a strong verifier. Cutting below the unit that carries its own meaning loses context, up to 27.6 points in one preprint. The benchmark is thus half the evidence for splitting. Claiming it for human reasoning needs that ceiling attached.\nstrength: medium\n\n### F12\ntarget: docs/research/whose-tokens-pay-for-mapping.md:551\ntarget_words: \"already carry a maturity stage, and public reading already holds back drafts.\"\nolder: docs/research/dogfood-run-4-agent-surface.md:57\nolder_words: \"Grep confirms **no other module uses it** — every working endpoint bypasses it.\"\nkind: narrows\nwhy: On 12 August the draft filter `public_claim_filter()` was found to be used by the feed alone. The feed had returned errors on every mode for four months because of that filter, and every working endpoint bypasses it. A grep today still finds only `feed.py` importing it. So \"public reading holds back drafts\" is true of the feed, not of claim pages or queries. Ledger X4 (drafts first) needs the filter wherever a person's map would be read.\nstrength: strong\n\n### F13\ntarget: docs/research/whose-tokens-pay-for-mapping.md:190\ntarget_words: \"Their acceptance of the map is the check that matters, and it is\"\nolder: docs/research/mapping-a-structured-position-on-ai-governance.md:46\nolder_words: \"claim he typed HYPOTHESIS and the map stored as `observed` was carrying **0.30 more confidence than\"\nkind: supplies-a-missing-piece\nwhy: A mapper rendered a friend's position with faithful passages, yet 20 claims he had typed as hypotheses or unresolved were stored as observed, carrying 0.30 more confidence than their author gave. Only the author reading the live map caught it. That supports making the author's acceptance the check that matters. It also shows the word-for-word and number checks cannot see a claim's epistemic weight, so a weight check must travel with them.\nstrength: medium\n\n### F14\ntarget: docs/research/deliberus-as-alignment-infrastructure.md:342\ntarget_words: \"Manipulated arguments that contradict the existing graph become visible through auto-connect edge conflicts.\"\nolder: docs/research/dogfood-run-6-israel-palestine-cross-domain.md:125\nolder_words: \"two claims that accurately report the same fact are classified as supporting each other at maximum confidence while their authors are in direct opposition.\"\nkind: contradicts\nwhy: Run 6 measured the relationship classifier returning SUPPORTS 1.00 for two renderings of one holding by authors in direct opposition. The 14 August correction says no such edge was stored then. A later note found 14 false SUPPORTS edges actually written between other pairs and retired by hand. Auto-connect has produced agreement across opposed material, not conflict edges. It cannot be counted on to expose injected contradictions.\nstrength: medium\n\n### F15\ntarget: docs/research/fragile-checkers-and-the-verification-bottleneck.md:95\ntarget_words: \"Cross-family agreement is the first confidence signal available to us that is behavioral, cheap, and calibratable.\"\nolder: docs/research/whose-tokens-pay-for-mapping.md:305\nolder_words: \"to all twelve links, so half its agreement is the share of Claude's answers that were *strong*.\"\nkind: narrows\nwhy: On the relevance judge's five-step scale, Claude agreed with its own earlier run on 10 of 12 links and three other models agreed with it on 4 to 6. Gemini 3 Flash answered \"strong\" to all twelve. Cross-family disagreement on degree questions partly records how each model uses the scale, not fragility. A two-family jury is clean on sharp yes/no questions, and its use for degree questions needs a shared answer form first.\nstrength: medium\n\n### F16\ntarget: docs/research/world-models-and-the-ontology-revival.md:55\ntarget_words: \"That is precisely what the founder described wanting: scheduled agentic workflows that continually structure, clean and understand the graph and its ontology.\"\nolder: docs/research/the-tidying-line.md:31\nolder_words: \"Any regular housekeeping should be done that could be done\"\nkind: already-answers\nwhy: On 2026-09-13 the founder ruled that regular housekeeping doable with high confidence in safety and accuracy should be done by machine. The same day he ruled that a machine-written link failing a confident re-check is replaced by itself, with the old link kept visible. For the graph the instinct is therefore ruled, not just \"unsolved in the field\". Open remain the schema itself and the risk that a tidy graph looks examined (ledger R11).\nstrength: medium\n\n### F17\ntarget: docs/research/sff-2026-and-fiscal-sponsorship-path.md:139\ntarget_words: \"Those framings are genuinely compatible with Deliberus's core thesis, but they are not the *entirety* of that thesis.\"\nolder: docs/research/convergence-wager-typed-residues-synthesis.md:14\nolder_words: \"2. **\\\"Instrument, not product\\\" is rejected** as a third successive shrinkage of the project's identity\"\nkind: already-answers\nwhy: On 5 July the founder rejected \"instrument, not product\" as the third successive shrinkage. Each earlier name (commercial product, personal tool, \"grant-shaped apparatus\") described the project by a smaller, more legible container. The note's mission-drift risk is that same pattern at a funder's door. The settled name, public epistemic infrastructure, gives the fixed destination. AI-safety, HSEE and Surplus framings then become entrances to it, not its description. A later outlook note adds that the pull is strongest exactly when funding arrives.\nstrength: medium\n\n### F18\ntarget: docs/research/jev-and-the-judgement-layer.md:429\ntarget_words: \"yes/no: *does this sentence weigh one consideration against another?*\"\nolder: docs/research/wild-weighing-dialects-and-the-question-set.md:11\nolder_words: \"Almost nothing in 72 wild sentences was \\\"X väger tyngre än Y.\\\" The forms that actually occur:\"\nkind: supplies-a-missing-piece\nwhy: The weighing question is not a crisp short yes/no in the wild. Blind-labelling the same 72 Swedish sentences gave ten hard cases and ten species of weighing (agenda, proportionality, urgency, betrayed-priority, enthymematic and others). The textbook \"X outweighs Y\" form almost never occurs, and four further questions were blessed and shipped for it. A Jev trial needs that species list as its question and the ten hard cases as its test, or it measures only the textbook form.\nstrength: medium\n\nNoticed: the founder's 2026-08-31 blessed end-state (no ratify act) is carried as a ⚠ pointer in `conviction-and-critique.md` § 1, but neither `grounded-critics-and-quarantined-imagination.md` nor `formalized-fermat-and-the-oracle-boundary.md` points the same way. Several of my nine are stale in places beyond the findings: `ai-safety-and-tao-augmentation-research.md` still cites the March counts (318 claims), and `sff-2026-and-fiscal-sponsorship-path.md` still cites \"60+ docs\". I also left unreported `blobs-community-sandboxes-and-the-advice-process.md`, whose membrane rule (nothing leaves by default) closes the Jev route for compartment material unless the membrane opens it. It is consistent with the Jev note, so it is not a conflict.\n"}