{"path":"research/corpus-graph/readers/2026-10-10-links-structure-and-agents-as-a-reader-returned-it.md","content":"# What a reader found to bear on eight notes on links, structure and agents as readers (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 8 notes:\n\n- `docs/research/how-far-a-premise-is-from-its-claim.md`\n- `docs/research/what-a-link-rests-on.md`\n- `docs/research/the-load-bearing-unsaid.md`\n- `docs/research/structure-versus-scale.md`\n- `docs/research/agents-as-a-consumer-class.md`\n- `docs/research/lean-deliberus-analogies.md`\n- `docs/technical-direction.md`\n- `docs/conceptual-threads.md`\n\nIt made 118 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\n8 of 8 notes read in full.\n\n### F1\ntarget: docs/conceptual-threads.md:273\ntarget_words: \"Each new extraction discovers connections to existing claims. The graph gets denser and more valuable over time.\"\nolder: docs/research/the-reuse-flywheel-prior.md:189\nolder_words: \"**The flywheel's shape as measured on the OLD linker**: per-source cross-source connection yield does not rise with\"\nkind: contradicts\nwhy: Measured 2026-08-29 on 31 sources, with the old cosine-only linker. A source's 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, and the note says this goes against the reading that it gets better as it grows. The sample is suggestive, not settled, but \"validated\" and \"denser over time\" are stronger than the record supports.\nstrength: strong\n\n### F2\ntarget: docs/research/structure-versus-scale.md:105\ntarget_words: \"Nothing currently measures it. GraphRAG-Bench's ten points on multi-hop is a proxy from a different task, not a test of this one.\"\nolder: docs/research/dogfood-run-7-swedish-election-structured-vs-unstructured.md:274\nolder_words: \"earn its cost\" as of this run: **the typed map does; the typed numbers are a promissory note.**\"\nkind: already-answers\nwhy: Run 7 (2026-08-17) is the test this passage says does not exist. Its control arm is prose plus a model, as the passage asks. The prose arm also found the crux. The typed map gave the crux an address someone can attack, and the typed numbers added nothing yet (every badge 0.500). It is one debate, hand-extracted. A reader weighing typed against cheap structure now has a first, mild answer, and the note never cites it.\nstrength: strong\n\n### F3\ntarget: docs/research/structure-versus-scale.md:123\ntarget_words: \"measured — 88% of concepts reused, a flatter-than-Zipfian slope, and the safety condition that reuse only transfers where sense is stable\"\nolder: docs/research/the-reuse-flywheel-prior.md:168\nolder_words: \"concept layer. But **only 4 concepts span more than one SOURCE (4%), and no concept in the graph spans\"\nkind: later-correction\nwhy: The 88% reuse figure is within-source. Measured 2026-08-29: 79 of 90 concepts are reused by more than one claim, but only 4 (4%) span more than one source and none spans more than two. So concept identity has the same matching problem as claim identity. The commons bet's \"evidence today\" does not yet show the cross-source reuse that coverage-scaled value needs.\nstrength: strong\n\n### F4\ntarget: docs/conceptual-threads.md:274\ntarget_words: \"The correction UX is live — answering CQs shifts QBAF badges.\"\nolder: docs/research/dogfood-run-7-swedish-election-structured-vs-unstructured.md:193\nolder_words: \"an unanswered critical question contributes zero by design, and **no CQ polarity claim in the corpus has ever received evidence**\"\nkind: narrows\nwhy: Measured 2026-08-17 on 250 random live claims. 116 carry evidence and 115 compute to exactly 0.500, because no critical-question polarity claim has ever received evidence. The correction UX is live, but the strength layer is \"correct and inert\": nobody has answered a question, so no badge has moved. \"Validated\" here means the code runs, not that answering shifts anything in practice.\nstrength: strong\n\n### F5\ntarget: docs/technical-direction.md:106\ntarget_words: \"Three primary axes from sketches: truthiness, relevance, controversiality\"\nolder: docs/research/relevance-and-the-two-factors.md:44\nolder_words: \"relevance living on the link, proposed by a machine, argued by people, and never voted.\"\nkind: already-answers\nwhy: Ruled 2026-09-28 (ledger Q33, founder: \"your proposal is blessed, make it so\"). Relevance is a weight on the link. A machine of a second model family proposes it, people argue it through the link's questions, and it is never voted. The Tier 3 line that lists relevance as one of three primary vote axes on a claim is out of date. It would mislead whoever builds the feed or the ballot.\nstrength: strong\n\n### F6\ntarget: docs/research/the-load-bearing-unsaid.md:65\ntarget_words: \"The brake currently rests on uninvited dissection being experienced as desecration.\"\nolder: docs/research/the-residual-error-taxonomy.md:164\nolder_words: \"analysis may go anywhere; what changes is *proportional attunement* and heightened humility, and only a premature verdict is prohibited\"\nkind: narrows\nwhy: Founder-ruled 2026-08-25, after this note: the sacredness brake no longer stops the descent. Analysis may go anywhere. At the boundary what changes is proportional attunement and humility, and only a premature verdict is prohibited; the felt datum alone is protected. So \"the wrong tool for a hinge\" and \"category-mistaken\" should be read as a duty of attunement and a limit on verdicts, not as a stop sign.\nstrength: strong\n\n### F7\ntarget: docs/technical-direction.md:198\ntarget_words: \"structured argumentation makes low-quality contributions structurally visible — a malformed argument with no supporting claims is self-evidently weak.\"\nolder: docs/research/steelmanned-critiques.md:167\nolder_words: \"This claim makes manipulation harder to detect because the manipulated output looks like rigorous reasoning rather than a biased vote count.\"\nkind: contradicts\nwhy: The critique argues the opposite. Because Deliberus claims to represent argument structure, a manipulator's output looks like rigorous reasoning, \"structurally invisible to ordinary quality control\". A sockpuppet network can rate premises \"well-argued\", and a poisoned chain looks tidy. A well-formed contribution is the threat, not a malformed one. Anyone lowering identity barriers because \"the structure is the filter\" should read it first.\nstrength: strong\n\n### F8\ntarget: docs/research/the-load-bearing-unsaid.md:35\ntarget_words: \"This kind of gap is **recoverable in principle**, which is why an entire NLP subfield exists to recover it\"\nolder: docs/research/cross-source-premise-run.md:80\nolder_words: \"Ten premises, zero junk, on four pairs, for under a dollar.\"\nkind: supplies-a-missing-piece\nwhy: Measured 2026-08-27, never cited here. The cross-source premise pass on four opposed pairs gave ten premises with zero generic, zero restatement and zero common-ground items. It recovered the \"occupation\" premise (54 uses on one side, none on the other). All ten were frame-level, not first-order, which is where this note says a crux shades into a hinge. Recoverability is shown for opposed pairs only, n=4, and no person has judged whether they bear load.\nstrength: strong\n\n### F9\ntarget: docs/conceptual-threads.md:190\ntarget_words: \"Trust must be EARNED through demonstrated epistemic competence, not declared through credentials.\"\nolder: docs/research/steelmanned-critiques.md:169\nolder_words: \"But it does not prevent *information environment* gaming: flooding the system with false evidence, creating authoritative-seeming sources\"\nkind: narrows\nwhy: The same critique says strictly proper scoring stops only gaming of declared belief. A well-resourced operation can be right on every uncontested question while manipulating the contested ones, so it earns any calibration score it wants. A scheme where a track record \"weights their future contributions\", as Thread 7 proposes, inherits that hole. The thread does not mention it.\nstrength: medium\n\n### F10\ntarget: docs/research/agents-as-a-consumer-class.md:130\ntarget_words: \"`REPORTS` is non-propagating only because those edges happen to carry no scheme.\"\nolder: docs/research/the-squirrel-and-the-strength-calculus/strength-rulings-and-code-as-read-2026-10-08.md:181\nolder_words: \"takes `(s)-[r]->(c)` where `r.scheme` is non-empty, `type(r)` is NOT in ['REPORTS','CLASSIFIES_TERMINUS','ARGUES_REQUIREMENT','SIMILAR_TO','ARGUES_RELEVANCE']\"\nkind: later-correction\nwhy: The second of the two named blockers was closed 2026-09-17. The badge query now names the non-propagating edge types and excludes REPORTS explicitly; TODO.md records it as done, and a code reading of 2026-10-08 (a helper's report) confirms the list. The first blocker, provenance labelling with count_safe_summary still uncalled, remains open. The gate for agent-facing shipping is one item shorter than § 5 says.\nstrength: medium\n\n### F11\ntarget: docs/research/agents-as-a-consumer-class.md:197\ntarget_words: \"(already open). It decides what a client may read as strength.\"\nolder: docs/ux-principles.md:380\nolder_words: \"The grey `UNEXAMINED` state is a restoration, not an invention.\"\nkind: later-correction\nwhy: The badge boundary was settled on the display side. A claim with edges and nothing answered now reads grey UNEXAMINED (founder-named 2026-08-17). A balanced one reads UNSETTLED (2026-08-23). A \"nothing stands\" state was ruled 2026-09-30 but is unbuilt. Item 1 of § 11 is largely answered. What is left is what the agent surface returns for such a claim, and the note could say that.\nstrength: medium\n\n### F12\ntarget: docs/technical-direction.md:158\ntarget_words: \"Already running on Darwin (darwin.home:6380). Graphiti's features map well:\"\nolder: docs/research/supersession-and-bitemporal-lifecycles.md:44\nolder_words: \"No integration is implied (the fleet's Graphiti instance serves session memory, a different job).\"\nkind: later-correction\nwhy: The stack table says the product graph is FalkorDB used directly. This note, written the day edge supersession shipped, says the fleet's Graphiti instance serves session memory and no integration is implied. Only the bitemporal shape is borrowed, as a model. The section headed \"FalkorDB via Graphiti\", with entities as claims, describes a choice that was not made and contradicts the table below it.\nstrength: medium\n\n### F13\ntarget: docs/research/how-far-a-premise-is-from-its-claim.md:206\ntarget_words: \"that the link is one step, or that it is a reason at all, and the no-fit answer cannot be used as a\"\nolder: docs/research/scheme-set-exhaustiveness.md:277\nolder_words: \"**The gate changes fewer answers than a second reading of the same question does.** 3 of 24 links\"\nkind: narrows\nwhy: § 3b and AD5 treat the classifier's no-fit exit as a menu item. The ruling of 2026-09-22 says a no-answer belongs in a gate. On 2026-10-10 the scheme step was gated (built, not deployed) and tried on 24 links. The gate moved 3 of 24, fewer than a second reading of the same question (4), and the scheme name itself was unsteady (stored name reproduced on 3 to 5 of 12). So 27 of 30 is a menu-form measurement, and the gate form has its own results.\nstrength: medium\n\n### F14\ntarget: docs/research/how-far-a-premise-is-from-its-claim.md:315\ntarget_words: \"**That the linking step's links are mostly wrong.** 19 of 30 are a reason at one step or one\"\nolder: docs/research/pass-quality-and-the-choice-of-model.md:111\nolder_words: \"links guessed between other people's claims, flags, proposals | Readers, as clutter. Undone by retiring it, the old link kept visible\"\nkind: supplies-a-missing-piece\nwhy: A tolerance table proposed 2026-10-05 (unruled, ledger Y1) allows 5 to 10 in 100 for machine-guessed links that move no strength and can be retired. By hand, 7 of 30 of the step's links were not reasons and 11 more needed a further claim. So \"not mostly wrong\" does not say whether the step sits inside any tolerance. The same note measured the step asked twice: about one link in seven would not be written the same way again.\nstrength: medium\n\n### F15\ntarget: docs/research/what-a-link-rests-on.md:316\ntarget_words: \"that comes back under several links, from as many separate readings, is a finding. It joins the pass that\"\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: same-idea-earlier\nwhy: The pass that reads a pair whole was run on four pairs on 2026-08-27 and was itself unsteady, the same unsteadiness this note measures for single namings. It also stored eleven proposal-grade premises and 50 edges in the live graph (machine links between texts). Recurrence across \"different routes\" therefore rests on a pass with a measured stability problem, and the 871-link debt in § 4.3 may leave those 50 edges out.\nstrength: medium\n\n### F16\ntarget: docs/research/lean-deliberus-analogies.md:204\ntarget_words: \"Deliberus needs temporal versioning of claims, definitions, and argument structures\"\nolder: docs/research/supersession-and-bitemporal-lifecycles.md:23\nolder_words: \"**deliberately does not handle this with lifecycle at all**: learning-better is an argument move — an ATTACKS edge,\"\nkind: later-correction\nwhy: Edge supersession shipped 2026-08-22. This note splits three things §4 runs together. The world changed: staleness, ratified by a person. The representation sharpened: supersession, naming its successor. \"We learned better\": deliberately not a lifecycle event but an argument move that stays contestable. \"Superseded by\" covers only the second. Temporal versioning is partly built (recorded_at, a staleness daemon) and partly refused by design.\nstrength: medium\n\n### F17\ntarget: docs/research/lean-deliberus-analogies.md:101\ntarget_words: \"Every premise the user didn't elaborate on gets a `sorry`. Every value they asserted without justification gets a\"\nolder: docs/research/threads-synthesis-and-gaps.md:107\nolder_words: \"**the completeness oracle publicly penalizes the exact state the sacredness brake protects.**\"\nkind: supplies-a-missing-piece\nwhy: Written 2026-07-10. The shipped sacredness brake keeps a person's own stated values from being decomposed uninvited, while the completeness oracle publicly flags unsupported value premises as the copout-axiom failure. A voice flow that marks every unjustified value as a sorry produces that clash: a flagged state the brake forbids repairing. It predates the 2026-08-25 re-scope of the brake, so check whether it still stands before building the flow.\nstrength: medium\n\n### F18\ntarget: docs/research/lean-deliberus-analogies.md:258\ntarget_words: \"the system can't re-validate the entire graph. It needs dependency tracking and incremental re-validation:\"\nolder: docs/research/synthesis-build-plan.md:168\nolder_words: \"The cache key carries a **subgraph fingerprint** — the retrieved claim ids with their strengths, plus the depth\"\nkind: same-mechanism-elsewhere\nwhy: Same mechanism, already built for syntheses. An artifact's cache key carries a fingerprint of the claims it read and their strengths, so it marks itself stale when its premises change, and a sweep marks drift as superseded. The note also records the limits: the cheap request-time check is blind to a strength moving, and recomputing the fingerprint per page view defeats the purpose. That is the practical price of Lake-style re-validation on a graph.\nstrength: medium\n\nNoticed: Two candidates I verified but did not report. steadying-an-outside-judge.md names the agents note and gives a cheaper test of what an agent reader gains from the graph (a fabricated-consensus lookup on the fact layer). the-group-premise-evidence.md line 45 counts \"independence before pooling\" as supplied by the graph, but agents-as-a-consumer-class.md:158 says agent readers sharing one graph are correlated by construction. The \"EXPLORATORY ... no architecture decisions\" banner at technical-direction.md:3 also sits beside a shipped-stack section. found.jsonl grew from 228 to 246 lines while I worked, and my 18 pairs were absent at the last check (the how-far to what-a-link-strength-means rating pair was already kept and is omitted).\n"}