{"path":"research/corpus-graph/readers/2026-10-10-reuse-sameness-and-machine-proofs-as-a-reader-returned-it.md","content":"# What a reader found to bear on eleven notes on reuse, sameness and machine proofs (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 11 notes:\n\n- `docs/research/the-reuse-flywheel-prior.md`\n- `docs/research/sameness-merge-and-split-across-fields.md`\n- `docs/research/the-canon-distillation-pass.md`\n- `docs/research/retrieval-instruments-beyond-cosine.md`\n- `docs/research/alien-proofs-and-the-library.md`\n- `docs/research/formalized-fermat-and-the-oracle-boundary.md`\n- `docs/research/lean-social-system-research.md`\n- `docs/research/bend-and-deliberus.md`\n- `docs/research/the-jhu-decomposition-line.md`\n- `docs/research/split-review-and-the-founders-intentions.md`\n- `docs/research/simplenote-archive-analysis.md`\n\nIt made 116 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\n11 of 11 notes read in full.\n\n### F1\ntarget: docs/research/formalized-fermat-and-the-oracle-boundary.md:84\ntarget_words: \"— so the same sentence here reads *N claims nobody examined*, which is the open founder question about\"\nolder: docs/research/the-ratification-record-and-the-triage-feed.md:210\nolder_words: \"challengeable forever; \"untouched\" becomes the permanent honest default rather than a debt awaiting\"\nkind: already-answers\nwhy: On 2026-08-31 the founder blessed the end-state \"there is no ratify act\". Machine discernments carry machine provenance forever, displayed as such and challengeable forever. Untouched structure is the permanent honest default, not a debt awaiting a gate. That answers the standing of untouched machine structure when no kernel exists. A reader should treat it as ruled, with only the display rule unbuilt, and not as an open founder question.\nstrength: strong\n\n### F2\ntarget: docs/research/split-review-and-the-founders-intentions.md:590\ntarget_words: \"Does rung three, sampled review with no panel, need a number he sets or one read off a\"\nolder: docs/research/a-cognitive-ecology-and-the-terms-of-the-relation.md:261\nolder_words: \"- A machine may propose, split, tidy and measure. Review need not keep pace with a machine pass\"\nkind: already-answers\nwhy: This restates ledger T1, ruled 2026-09-29 with the founder's \"Yes.\" and admitted as a canon line. A pass with a measured error rate keeps the same share of wrong content at any volume. Only machine-on-machine depth compounds, so the measured rate caps depth. With no measured rate a person engages first. That answers the question here: the number is the measured rate against his tolerance, read off a sitting, and a per-claim panel is not what keeps volume safe.\nstrength: strong\n\n### F3\ntarget: docs/research/lean-social-system-research.md:270\ntarget_words: \"Different roles have different authority — contributors propose, reviewers validate, maintainers adjudicate\"\nolder: docs/research/the-ratification-record-and-the-triage-feed.md:179\nolder_words: \"The § 3b demotion, carried to its end: assent was already telemetry —\"\nkind: contradicts\nwhy: Blessed 2026-08-31 (\"blessing this whole proposal\"): the graph holds contributions and telemetry only, and verdict-acts do not exist. Assent is telemetry, and dissent must arrive as an argued reclassification. The only carve-out is operator-phase review by the founder and friends. Copying Mathlib's contributor, reviewer, maintainer ladder into Deliberus would build the apparatus he rejected. His 2011 validator ladder was the same apparatus.\nstrength: strong\n\n### F4\ntarget: docs/research/sameness-merge-and-split-across-fields.md:107\ntarget_words: \"do not build \"ask the model whether these two claims are the same\" as the core\"\nolder: docs/research/auto-connect-upgrade.md:30\nolder_words: \"the pairwise seven-key check runs *first* on every candidate pair\"\nkind: contradicts\nwhy: The founder ruled on 2026-09-13 (\"Confirmed the proposed ordering\") that a one-call model seven-key check judges every candidate pair before any relationship is classified. It was built 2026-09-22 as keeper.py. Session 36 found the older classifier re-asking sameness, so the less careful asker could overrule the keeper. The note's advice to normalise each claim and let a dedicated matcher decide was not the path taken. Anyone acting on it should read the ruling first.\nstrength: strong\n\n### F5\ntarget: docs/research/the-jhu-decomposition-line.md:122\ntarget_words: \"was written into the graph: a part becomes a claim only when a person keeps it.\"\nolder: docs/research/split-review-and-the-founders-intentions.md:298\nolder_words: \"Then: *\"Wire and switch first, sitting later.\"* The operator sweep then\"\nkind: later-correction\nwhy: This section says nothing was written and rejects storing parts as settled. The same evening the founder answered \"Wire and switch first, sitting later\". The splitter went from proposing to writing marked parts, and the sweep wrote 7 proposals and 44 parts on his \"Oh, sure, do it\". Ledger B7 now reads ruled and overtaken. The note half-concedes this (\"writing-marked\", \"B7, open\") yet keeps the proposal-only account. The current state is the split-review § 1d.\nstrength: strong\n\n### F6\ntarget: docs/research/the-jhu-decomposition-line.md:24\ntarget_words: \"All 131 decomposition edges are untagged, so it protects nothing today.\"\nolder: docs/research/session26-open-rulings.md:45\nolder_words: \"padding pinned by regression tests — 3 vs 12 untagged children now score identically\"\nkind: later-correction\nwhy: Ledger B10 records the founder's ruling of 2026-08-31, \"Go with option C, both\". The default flip went live that day, so untagged decomposition children are treated as necessary. Roughly 80 parents are now governed by their weakest part. The \"inert defence, tag the edges\" step in this section is overtaken, and the splitting-inflation attack is already closed by default. The founder's confirmation of its two felt consequences is still awaited.\nstrength: strong\n\n### F7\ntarget: docs/research/simplenote-archive-analysis.md:194\ntarget_words: \"Deliberus as the next evolutionary step in collaboration, following multicellular life → culture → civilization.\"\nolder: docs/research/major-transitions-heredity-and-evolvability.md:445\nolder_words: \"civilisation, can stand under this reading: a new level of organisation built on difference, with no\"\nkind: narrows\nwhy: The note says this frame is \"not in any existing doc\". It is now tested at length against the major-transitions literature. The founder ruled on 2026-10-10 (ledger S8, \"Yes!\") that members give up working everything out alone but keep their own position, the ability to check and the freedom to leave. The 2012 sentence survives only as a new level of organisation built on difference, with no claim that a new individual formed. A hive-mind reading is excluded.\nstrength: strong\n\n### F8\ntarget: docs/research/simplenote-archive-analysis.md:231\ntarget_words: \"- Promotion is **automated by the system** based on vote-accuracy metrics\"\nolder: docs/research/split-review-and-the-founders-intentions.md:326\nolder_words: \"In 2011, review was queues, validators and a *\"marked as valid\"* state. On 2026-08-31 he blessed\"\nkind: later-correction\nwhy: The 2011 validator ladder, with its queues and promotion by vote-accuracy, was reversed: \"there is no ratify act\", and bare judgments are telemetry. The narrower survival is that whoever judges earns the right by measured agreement. That now applies to machine passes through the trust ladder (proposed, unruled), not to ranked users. Whoever mines these notes for a user-level or validator design should read this reversal first.\nstrength: strong\n\n### F9\ntarget: docs/research/retrieval-instruments-beyond-cosine.md:112\ntarget_words: \"| **Concepts** | Already works, 87% reuse | Cosine plus the existing concept layer |\"\nolder: docs/research/the-reuse-flywheel-prior.md:168\nolder_words: \"But **only 4 concepts span more than one SOURCE (4%), and no concept in the graph spans\"\nkind: later-correction\nwhy: The 87% concept reuse is a within-source figure. Measured 2026-08-29, only 4 of 90 concepts span more than one source, and none spans more than two. Concept labels are minted per extraction with source-specific phrasing, so concepts have the same matching problem as claims. The table row that calls concepts \"already works\" and leaves them to cosine plus the concept layer is therefore not safe for cross-source use.\nstrength: strong\n\n### F10\ntarget: docs/research/retrieval-instruments-beyond-cosine.md:109\ntarget_words: \"**Structural identity by construction** — same template, same slots.\"\nolder: docs/research/latent-scaffolding-and-the-type-token-split.md:24\nolder_words: \"if a claim's identity is structural — (scheme, CQ id, target) — then its text is\"\nkind: supplies-a-missing-piece\nwhy: The founder carried this ruling to its storage conclusion and endorsed it the same day: \"the compression and latent-until-touched sounds excellent, wanna implement that quite soon\". 3,110 stored scaffolding claims (64% of the graph) are tokens of a few templates. Store the type once and derive the text. The row only said to skip the embedding. The later note shows what structural identity means for storage, and that this population needs no merge pass at all.\nstrength: strong\n\n### F11\ntarget: docs/research/the-jhu-decomposition-line.md:39\ntarget_words: \"So the paper the pipeline is named after has an architecture Deliberus cannot express: the graph keeps the augmented form only.\"\nolder: docs/research/store-the-uses-read-the-meaning.md:35\nolder_words: \"paraphrase. And a claim's wording is a machine's rendering of a passage, stored as if it were the\"\nkind: supplies-a-missing-piece\nwhy: The note asks whether to store the pre-decontextualisation atom, and which part a user attacks. Arc AA (ledger AA3 and AA6, open) answers differently. Store a pointer to the author's passage, computable for 923 of 1,144 claims, with the wording as a dated rewrite. Measured: 29% of a claim's content words appear nowhere in its source, against 5 to 6% in hand extractions. The inserted material is more than context, and the author's passage is the cheaper thing to keep.\nstrength: strong\n\n### F12\ntarget: docs/research/the-canon-distillation-pass.md:154\ntarget_words: \"Until then a finding reached a session only by being resident or by being looked up.\"\nolder: docs/research/what-agent-memory-research-says-about-agent-surfaces.md:25\nolder_words: \"least five of them were lines in the instructions it had loaded. The note holding the ruling behind the\"\nkind: supplies-a-missing-piece\nwhy: That note reports, second-hand, that an agent with a memory store made no memory call in 114 turns. The same items delivered by the harness at set moments arrived 138 of 138 times. In this project a session met nine things already ruled, and at least five were resident lines. So residency also fails. Delivery at the moment of use is a measured third route, and the hook is that route, not only a convenience. This bears on the unruled loop.\nstrength: medium\n\n### F13\ntarget: docs/research/the-reuse-flywheel-prior.md:332\ntarget_words: \"predictable — a handful of debates per topic plausibly exhausts most of it, and the marginal value of\"\nolder: docs/research/saturation-and-the-long-tail.md:205\nolder_words: \"detector already exists: the `does_not_fit` rate. **A topic showing high estimated coverage AND a\"\nkind: narrows\nwhy: Written the same day, the saturation sweep splits the curve in two. The head closes (Chao-shaped), but the tail decelerates without ever closing (Heaps-shaped). Duplicate-detection recall falls as a corpus grows, so sameness gets harder exactly when it matters most. A saturating curve cannot tell a mapped domain from an ossified frame, and high coverage with a near-zero does-not-fit rate is \"suspicious, not finished\". \"The coverage curve is the budget\" needs that second dial.\nstrength: medium\n\n### F14\ntarget: docs/research/formalized-fermat-and-the-oracle-boundary.md:75\ntarget_words: \"silently in both domains, and in both the remedy is a second reader of the statement rather than a\"\nolder: docs/research/quality-control-for-deeper-steps.md:129\nolder_words: \"1. **The splitter's separate check is separate, and it is not blind.** It is one call. Its prompt\"\nkind: narrows\nwhy: The \"comparator by another name\" claim assumes the coverage check is an independent second reader. Read in the code on 2026-10-10: it is one call whose prompt prints the original claim beside the instruction to rebuild blind. Microtasks-in-broad-daylight § 3b measured that the check flagged none of six harmful parts until it was rewritten on 2026-09-30. Lean's comparator compares two formal statements. Ours is a model with the answer in view.\nstrength: medium\n\n### F15\ntarget: docs/research/bend-and-deliberus.md:184\ntarget_words: \"proper name absent from its parent. That check is deterministic, cheap, and has a known blind spot\"\nolder: docs/research/microtasks-in-broad-daylight.md:160\nolder_words: \"see it: nothing was dropped and nothing was added, since every word came from the sentence, with its\"\nkind: narrows\nwhy: Bend proposes a deterministic law that a part may not contain a number, date or name absent from its parent. It concedes a blind spot for invention using no new ones. Microtasks measured where the harm actually lives. Of 37 written parts, six stated as fact a clause the claim denies or concedes. These are polarity flips built from the sentence's own words, and two came back coverage-clean. The law would not see them, so the \"known blind spot\" is where the measured harm is.\nstrength: medium\n\n### F16\ntarget: docs/research/lean-social-system-research.md:182\ntarget_words: \"- Should be confluent (order of application shouldn't matter)\"\nolder: docs/research/bend-and-deliberus.md:135\nolder_words: \"So for 59% of the pairs, whether auto-connect ever judges them depends on which of the two sources\"\nkind: same-mechanism-elsewhere\nwhy: Lean enforces confluence on its growing rule set by review and linters. Bend measured the same property failing in Deliberus's own graph builder. For 59% of 1,729 candidate pairs, whether auto-connect ever judges them depends on which of two sources arrived second (ledger T4, open). The Lean note presents the \"growing competence\" flywheel without any such condition. It is evidence that order-independence is a requirement the build lacks, not a nicety.\nstrength: medium\n\n### F17\ntarget: docs/research/alien-proofs-and-the-library.md:329\ntarget_words: \"first step is a count that has not been made: how many links in the stored sources have a rule\"\nolder: docs/research/quality-control-for-deeper-steps.md:280\nolder_words: \"| Chains an author wrote: a first claim supports a middle claim, which supports a conclusion, all in one text\"\nkind: supplies-a-missing-piece\nwhy: AF4 proposes hiding something an author wrote and counting recovery, and says the first step is a count not yet made. The sibling note made the matching count for its own version of the test. 200 chains of two steps in 20 texts and 85 of three can have the middle hidden. 1,789 claims carry their own passage. It is the same design for middle steps rather than rules, and it shows how concentrated the answer keys are: 70, 34 and 29 chains sit in three texts.\nstrength: medium\n\n### F18\ntarget: docs/research/alien-proofs-and-the-library.md:179\ntarget_words: \"for two errors this corpus already tracks. Being paid for shortness is being paid for leaving\"\nolder: docs/research/the-canon-distillation-pass.md:130\nolder_words: \"a topic entry from 1,636 to 479 tokens, dropping 24 of 25 sentence-units. Audited on the founder's\"\nkind: same-mechanism-elsewhere\nwhy: This is the project's own measurement of a shortening pass. A size-driven trim of a resident entry (1,636 to 479 tokens) dropped 24 of 25 units. On audit, two of the dropped units passed the keep-filter and were restored. The pass's lesson: \"trimming by size is not trimming by value\". That confirms the claim that a length score pays for leaving things out and cannot tell pointing from omitting. It also shows the failure on a human-run compression pass, not only a machine one.\nstrength: medium\n\nNoticed: The split-review note names the ledger, SESSION-HISTORY.md and pass-quality-and-the-choice-of-model.md. I therefore could not use the founder's 2026-10-05 ruling on the live splitter (ledger Y5: \"Let's just have me go through and label/review quite soon, we don't have to hold\") against it. That ruling bears on its § 5 questions 1 to 3. The JHU note's top half (lines 19–60) is stale in several places, though its later sections correct some of it. CLAUDE.md is now about 413 KB and the global file about 366 KB, against the canon note's 100,599 and 80,226 tokens measured 2026-08-29, so the budget section's figures have drifted.\n"}