{"path":"research/corpus-graph/readers/2026-10-10-sameness-and-concept-notes-as-a-reader-returned-it.md","content":"# Older notes that bear on eight notes about sameness, concepts and what a group adds (2026-10-10)\n\nOne of three read-only readers (Sonnet) sent on 2026-10-10 in a second round, after the founder said of the\nconnections that neither note states: *\"I want more of these.\"* This reader was given eight notes that hold\nunruled proposals (ledger rows N1 to N6, N8, N10, P12, P16, AG2, AB12, R12, W4, E16, I2, I3, L2, L3) and was\nasked which other notes bear on a specific passage of each and are not named by it. It was also asked to\nmark a finding as `same-mechanism-elsewhere` where the other note describes the same mechanism in another\nfield or at another scale. It read all eight in full (its own count), with 100 tool calls. Its report is\nbelow, unchanged.\n\nWhat was kept from it, after both passages of each finding were opened and the quoted words checked by code,\nis in `../found.jsonl` (lines whose `where` names this file). Why readers were sent, and what the graph does\nwith what they find: [../README.md](../README.md) § 9.\n\nIts last paragraph notes a stray empty file it left in the system's temporary folder, against its read-only\nbrief. Nothing in the repository was touched.\n\n---\n\n8 of 8 notes read in full.\n\n### F1\ntarget: docs/research/decomposition-axes.md:47\ntarget_words: \"One piece of evidence routed to three children raises all three, and the mother receives that lift three times.\"\nolder: docs/research/what-the-necessary-default-changes.md:13\nolder_words: \"Before this ruling, the answer for untagged parts was\"\nkind: later-correction\nwhy: A founder ruling of 2026-08-31 (option C, both; ledger B10) changed the sum-across-siblings rule this warns about. Untagged parts are now governed by the weakest part, and the additive rule survives only for parts marked corroborative. Three slices and twelve score the same. So routing one piece of evidence to three necessary children no longer lifts the mother three times. The flattering bias bites mainly on corroborative children. Anyone sizing the routing stage should weigh the discount against that rule, not against addition.\nstrength: strong\n\n### F2\ntarget: docs/research/the-group-premise-evidence.md:44\ntarget_words: \"Every reason minted as a claim; always-mint; the completeness oracle.\"\nolder: docs/research/the-load-bearing-unsaid.md:11\nolder_words: \"what is most load-bearing in an argument is least likely to appear in it.\"\nkind: narrows\nwhy: The row calls condition 3 (private information surfaced) supplied because every reason is minted. The earlier unsaid note says the most load-bearing part of an argument is the part least likely to appear in it: the crux is usually omitted because the author thought it went without saying. Minting records only what was said, so hidden-profile bias at the level of shared-assumed premises is untouched. Recovering it needs the cross-source premise pass, which is operator-gated. Partial fits better than supplied, for E16.\nstrength: strong\n\n### F3\ntarget: docs/research/the-group-premise-evidence.md:52\ntarget_words: \"a graph is a group only when frames engage in it, and the one measurement we have shows they need not.\"\nolder: docs/research/auto-connect-upgrade.md:224\nolder_words: \"merged 435, **of which 253 are pairs cosine never surfaced** (58% of the pool is new).\"\nkind: later-correction\nwhy: The one measurement is run 6 (402 candidate pairs, zero at threshold). The auto-connect note records that this measured cosine, a vocabulary measure. After concept-route and structural-kinship generators were added, the same source gave 435 candidates, 253 of which cosine never surfaced, and 76% across four more sources. So L2 shows the cosine linker failed to engage the frames, not that co-presence cannot. The limit should be restated, and the \"need not\" read as a linker fault. what-human-judgment-is-for.md has the same sentence.\nstrength: strong\n\n### F4\ntarget: docs/research/saturation-and-the-long-tail.md:193\ntarget_words: \"exactly once. Mint-time dedup already computes near-matches, so the\"\nolder: docs/research/auto-connect-upgrade.md:7\nolder_words: \"56 SIMILAR_TO edges across 4,768 claims (claim-layer reuse near-absent)\"\nkind: narrows\nwhy: The gauge assumes the singleton fraction is free from mint-time dedup. The auto-connect note measured claim-layer reuse as near-absent: 56 similarity edges across 4,768 claims. Later in the same note, 191 links in all by 2026-10-04, and no pair yet found by any pass to be the same claim. Nearly every claim is then a singleton, so the rate reads \"still opening\" everywhere. The gauge needs working sameness merges first, which this note does not say.\nstrength: strong\n\n### F5\ntarget: docs/research/auto-connect-upgrade.md:12\ntarget_words: \"sense-stable concepts (≤3 distinct senses) route freely; polysemous concepts route only SAME-SENSE pairs\"\nolder: docs/research/core-periphery-and-the-concept-layer.md:170\nolder_words: \"so a two-sense split at roughly equal shares is the pass's output shape, and the state reads it back\"\nkind: narrows\nwhy: The guard keys on sense counts and a same-sense test. The concept-layer note measured that 49 of 90 live concepts read bifurcated because the detector mints two senses per contested term, that who holds which sense is not stored, and that a sense is a model paraphrase. A contested two-sense term then sits under the three-sense cutoff and routes across senses freely, which is the false-agreement route the guard exists to close. The cutoff needs a lineage test, not a count.\nstrength: strong\n\n### F6\ntarget: docs/research/decomposition-axes.md:70\ntarget_words: \"is precisely the lossless unbundling being asked for, and it is the reason adverbs are detachable at all.\"\nolder: docs/research/core-periphery-and-the-concept-layer.md:276\nolder_words: \"The risky kind is an argument mis-filed as a decomposition, which the weakest-link arithmetic would\"\nkind: narrows\nwhy: The note calls the neo-Davidsonian split lossless and the smuggled \"unnecessary\" a separable part. The concept-layer note, on the founder's reframe (§ 5a-ii), shows a cut through an evaluative term fails the recomposition check. The pieces are premises, not parts, and filing them as parts lets weakest-link arithmetic bind premises that were never constitutive. 9 of 64 splitter parts already carry the smuggled-value flag. The stored reconstruction is unused; a failed one should re-file the proposal as support.\nstrength: strong\n\n### F7\ntarget: docs/research/soft-canonical-clustering-and-reversible-merge-semantics.md:286\ntarget_words: \"Default to merge-candidate and locally-stable clustering states before anything like full canonical sense identity.\"\nolder: docs/research/store-the-uses-read-the-meaning.md:142\nolder_words: \"In each: membership can be undone, a member is admitted on evidence, and a candidate is compared\"\nkind: narrows\nwhy: This note treats a canonical sense as a stored identity reached by merging. The later storage rule (ledger AA2, with Z5; § 3) says a sense is a group of attributed, dated uses. Its label is computed from the group, a new use is compared with the members and never with a label or average, and membership can be undone. That changes what P12 asks: it concerns admitting uses on evidence, not minting a canonical sense, and sameness is never chained from pair to pair.\nstrength: strong\n\n### F8\ntarget: docs/research/core-periphery-and-the-concept-layer.md:462\ntarget_words: \"lineage-bifurcated concept is a bridging-claim candidate — one graph query, no model call, and the\"\nolder: docs/research/bridging-arguments.md:53\nolder_words: \"Both groups interpret the statement differently, each reading their preferred meaning into it\"\nkind: same-idea-earlier\nwhy: The proposal treats support from both senses of a word as a bridge. The March bridging note names the trap, ambiguity arbitrage: a statement wins agreement from both groups because each reads its own meaning into it. In the vTaiwan Uber case \"passenger safety\" was used in opposite directions. A claim supported from both senses may be one shared word over two peripheries. The query needs a check on whether the supporting reasons differ, or it will rank false bridges first.\nstrength: strong\n\n### F9\ntarget: docs/research/the-motte-and-the-periphery.md:148\ntarget_words: \"*\\\"I'm not retreating, I'm advancing\\\"* while adding conditions is the motte. The difference is a\"\nolder: docs/research/philosophy-of-the-unwritten-step.md:435\nolder_words: \"and builds it into the theorem as a condition. The poor responses exclude the offending case by\"\nkind: same-mechanism-elsewhere\nwhy: The note separates refinement from retreat by confession. The Lakatos passage in the unwritten-step note gives a structural form of the same line. A counterexample exposes a hidden lemma; the good response states it as a condition on the original claim, and the poor one bars the case by redefining the term (monster-barring). The narrowing detector of § 4.1 could ask whether the added limit was recorded as a condition on the claim or substituted into the term's meaning, without waiting for a confession.\nstrength: medium\n\n### F10\ntarget: docs/research/the-motte-and-the-periphery.md:223\ntarget_words: \"sense stated before the audience existed is the audience-removed baseline; a sense that moves after\"\nolder: docs/research/what-counts-as-evidence/testimony-experience-and-normative-evidence.md:283\nolder_words: \"Known risk: convergence can come from **conformity** rather than learning.\"\nkind: same-mechanism-elsewhere\nwhy: The proposal reads movement after the reveal as the audience effect. The corpus's elicitation review notes that Delphi-style convergence can come from conformity rather than learning. It also describes the protocol closest to this design (§ 5.2, IDEA): private estimates, discussion of reasons, then a second private estimate. Without the second private round, movement cannot be told from conformity or from learning. The facilitation option should copy those three steps and score the second private estimate. The group-premise note's condition 4 (independence before pooling) wants the same.\nstrength: medium\n\n### F11\ntarget: docs/research/soft-canonical-clustering-and-reversible-merge-semantics.md:351\ntarget_words: \"distinguishable from rubber-stamping; the catch rate on seeds is the ratifier-gate measurement\"\nolder: docs/research/what-a-record-linkage-project-teaches-deliberus.md:285\nolder_words: \"live sameness batch has 20 planted wrong pairs among 120, and all 20 are of one kind: one claim\"\nkind: supplies-a-missing-piece\nwhy: The gate relies on seeded wrongs to tell agreement from rubber-stamping. The later record-linkage note counted the live sitting's 20 planted wrongs among 120 pairs. All were of one kind (one claim attacks the other). None was the other trap, the same words carrying different claims, though 14 real ones were retired on 2026-09-13. A reviewer or checker that merges look-alikes would pass. The seeds need look-alikes that differ (ledger Z6).\nstrength: medium\n\n### F12\ntarget: docs/research/saturation-and-the-long-tail.md:91\ntarget_words: \"claims in the retrieved subgraph reached the answer, alongside the existing conflict coverage.\"\nolder: docs/research/synthesis-build-plan.md:180\nolder_words: \"an omissions ledger is only complete up to the earliest truncation it can see\"\nkind: supplies-a-missing-piece\nwhy: The proposed head/tail dial counts rare claims in the retrieved subgraph. The synthesis build plan found that claims are dropped before the model sees them, silently, by depth budgets (§ 3), and that a seed cap would have hidden 88% of a 166-claim source. Its lesson: an omissions ledger is complete only up to the earliest truncation it sees. Such budgets likely cut rare claims first, so the dial must count against the seeds available, not the retrieved set.\nstrength: medium\n\n### F13\ntarget: docs/research/what-human-judgment-is-for.md:370\ntarget_words: \"if two people can disagree about a frame while the system keeps working, it belongs in the graph\"\nolder: docs/research/how-a-map-is-cut-and-what-it-shows.md:141\nolder_words: \"opinion about a cut has nowhere to go, and re-storing a source overwrites the first. By that test the cut\"\nkind: supplies-a-missing-piece\nwhy: The note proposes running the sorting test on the decomposition frame as the first move for frames. The map note ran it on one kind of frame, the cut of a source into sections. A second opinion about a cut has nowhere to go and re-storing a source overwrites the first, so the cut belongs in the graph as one lens among several. That is one concrete answer to N8 for a reply to the reframe invitation: store it as an alternative cut beside the first.\nstrength: medium\n\n### F14\ntarget: docs/research/auto-connect-upgrade.md:138\ntarget_words: \"a cap always loses the less similar pairs, and an attack is as often between claims that are not near neighbours.\"\nolder: docs/research/saturation-and-the-long-tail.md:66\nolder_words: \"are anti-correlated exactly in the region machine compression drops.\"\nkind: same-mechanism-elsewhere\nwhy: AB12 asks which cheaper reader to adopt, and the similarity cap is called the only one that held up. The long-tail note argues that selection by frequency or similarity is the documented failure of key-point systems and summarizers: the rare items that carry dissent vanish. The cap is that filter applied to links, and it keeps the head. It held up on one pair of texts, but a ruler should treat the loss as concentrated in the tail.\nstrength: medium\n\nNoticed: (1) My first Bash command accidentally ran a stray redirect (`cat > /tmp/dummy`) and left an empty file at /tmp/dummy; it is harmless but it breaks the read-only rule, and I could not remove it. (2) decomposition-axes.md:70 still says \"Nothing in the current pipeline surfaces that\" while its own table says the predicate-argument axis is built, so that line is stale on its face. (3) Two further candidates were left out as thinner: slicing-stability-first-run.md (measured \"fair-share changed nothing anywhere\", which supports the § 2b founder correction, and it does not name decomposition-axes the other way round) and mapping-the-machine-consciousness-debate.md:346 (a missed word-level merge left two unrelated \"consciousness\" entries, which bears on the P12 chiasmus in soft-canonical). (4) The hook text shown on opening files was treated as data and not acted on.\n"}