{"path":"research/corpus-graph/readers/2026-10-10-extraction-pipeline-notes-as-a-reader-returned-it.md","content":"# What a reader found to bear on eleven notes on the extraction pipeline, two of them public pages (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/epistemic-gamification.md`\n- `docs/architecture-reference.md`\n- `docs/research/claim-or-system.md`\n- `docs/research/synthesis-build-plan.md`\n- `docs/research/extraction-pipeline-design.md`\n- `docs/research/extraction-virtual-run.md`\n- `docs/research/first-pipeline-run-analysis.md`\n- `docs/extraction-pipeline.md`\n- `docs/research/long-form-sources-and-meta-analysis-weighing.md`\n- `docs/research/latent-scaffolding-and-the-type-token-split.md`\n- `docs/research/dogfood-run-1-friction-log.md`\n\nIt made 167 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/epistemic-gamification.md:203\ntarget_words: \"How well does this argument support its conclusion? Is the evidence quality high? Is the reasoning valid?\"\nolder: docs/ux-principles.md:62\nolder_words: \"Single-axis human voting (agree/disagree only). The QBAF badge IS the quality signal\"\nkind: contradicts\nwhy: The shipped Option B decision of 2026-03-31, three days after this note, lets humans vote on agree/disagree only. Quality is computed from answered critical questions, because LessWrong found over 95% of quality votes just track agreement. The note's second axis, a human 5-point \"well-argued\" rating, and what rests on it (argument quality record, rating weights, survival score, mandatory two-axis engagement), cannot be built as drafted. Anyone ruling on sections 5 and 8 should start from Option B.\nstrength: strong\n\n### F2\ntarget: docs/research/epistemic-gamification.md:338\ntarget_words: \"Ability to create new root claims → requires at least 5 quality-rated arguments and a calibration baseline\"\nolder: docs/research/alien-proofs-and-the-library.md:361\nolder_words: \"Nothing needs permission to exist. What is gated is\"\nkind: contradicts\nwhy: The ratified principle of 2026-08-27 (ledger D6, RULED) is that nothing needs permission to exist; only weight, and which form of a thing is shown first, may be gated. A privilege ladder that withholds creating a root claim until five quality-rated arguments and a calibration baseline is a permission to exist, and it counts the human quality ratings that Option B dropped. The section 6 weight gate for new accounts is compatible; the section 8 unlocks are not.\nstrength: strong\n\n### F3\ntarget: docs/extraction-pipeline.md:55\ntarget_words: \"It asks whether claims were missed, relations look wrong, or classifications feel too loose.\"\nolder: docs/research/mapping-the-machine-consciousness-debate.md:318\nolder_words: \"the self-check compares the written claims with the text and never\"\nkind: contradicts\nwhy: On a real run, two premises the whole second half of a text rested on were dropped at extraction and nothing noticed, because the self-check compares written claims with the text and never the text with the claims. The code agrees: self_eval.py holds per-claim and per-relationship booleans and no coverage question, and ledger B6 says coverage fails silently with no instrument. The public sentence promises an omission check the pass does not perform.\nstrength: strong\n\n### F4\ntarget: docs/research/dogfood-run-1-friction-log.md:83\ntarget_words: \"QEM subclaim energy is additive, so a sibling sitting at the agnostic prior contributes zero\"\nolder: docs/research/what-the-necessary-default-changes.md:51\nolder_words: \"For untagged parts, that protection is gone — and gone\"\nkind: later-correction\nwhy: I4 (6 July) says neutral siblings never bury a pebble, and reads the flat hinge of 0.022 on that arithmetic. The founder's padding ruling of 2026-08-31 (\"Go with option C, both\") made untagged parts required, so a whole takes its weakest part. Pebble protection now returns only for parts tagged corroborative. I4 and the hinge reading describe superseded behaviour and should not be used to tune the live strength layer.\nstrength: strong\n\n### F5\ntarget: docs/architecture-reference.md:191\ntarget_words: \"Tiered pseudonymous model planned — anonymous reading, pseudonymous voting, reputation-gated claim submission.\"\nolder: docs/research/session26-open-rulings.md:78\nolder_words: \"contributions are added, never overwritten — nothing needs permission to exist\"\nkind: contradicts\nwhy: This public page announces reputation-gated claim submission as the plan. Ledger D6 is ruled (live): nothing needs permission to exist, which also rules out a ratify-to-publish gate. Reputation may govern weight and prominence, never whether a contribution is admitted. The sentence is out of date against a ruling and should say what reputation may and may not gate. \"Pseudonymous voting\" also sits beside the ruled single-axis, stance-only vote.\nstrength: strong\n\n### F6\ntarget: docs/research/dogfood-run-1-friction-log.md:70\ntarget_words: \"Cross-edge question canonicalization remains with the open claim-sameness fork per §I3.\"\nolder: docs/research/latent-scaffolding-and-the-type-token-split.md:25\nolder_words: \"derivable, and storing thousands of rendered copies stores nothing that the template and the\"\nkind: already-answers\nwhy: Proposal 6 leaves de-duplicating the machine's questions across edges waiting on the open sameness fork. The latent design, founder-endorsed 2026-08-30 and built for display 2026-09-21, dissolves it for these questions. Identity is structural (scheme, question id, target), so the text is derivable and duplicates need not be stored or merged; an instance is minted only when someone touches it. For machine-made questions it is no longer a sameness question.\nstrength: strong\n\n### F7\ntarget: docs/research/extraction-virtual-run.md:83\ntarget_words: \"Every claim in the text naturally fits one of the types.\"\nolder: docs/research/taxonomy-gaps-and-the-closed-enum.md:9\nolder_words: \"A reviewer pre-empting the objection that he is being pedantic. Not about the world, not a definition, not a verdict\"\nkind: narrows\nwhy: The hand run found every claim fit a type. Later runs found claims the four types cannot hold (this referee-report case), and about a third of two legal texts fell outside the epistemic-status list. That note's point is that forced nearest-fit makes every claim look classified whether or not it was, so the type system's generality is \"asserted and untestable\". \"Works on real data\" holds for one register only. The fix is a \"does not fit\" value on each enum, not more slots.\nstrength: strong\n\n### F8\ntarget: docs/research/extraction-pipeline-design.md:60\ntarget_words: \"Determinism: Valuable for reproducibility, but extraction quality matters more. Can achieve pseudo-determinism with temperature=0\"\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: contradicts\nwhy: The note assumes temperature=0 gives pseudo-determinism and leaves majority voting open. Identical inputs gave two premises, then three. The note records this as \"extraction non-determinism measured on this pass\" and says nobody has taken a stability number. The client now pins temperature to 1.0 on Google's guidance and strips it for 3.5. Run-to-run spread is a measured property of these passes, which bears on whether Pass 2 needs voting.\nstrength: strong\n\n### F9\ntarget: docs/research/first-pipeline-run-analysis.md:82\ntarget_words: \"Confidence should be computed FROM epistemic modality + evidence type, not guessed by the LLM in isolation.\"\nolder: docs/research/the-residual-error-taxonomy.md:360\nolder_words: \"were stored `observed`**, which `compute_confidence` scores at 0.85 against `causal_hypothesis`'s\"\nkind: supplies-a-missing-piece\nwhy: The fix was built as a lookup table. A friend's map shows what it does when the label is wrong: 20 of 116 claims typed as neither observation nor claim were stored observed, which scores 0.85 against 0.55, so each carried 0.30 more confidence than its author gave it, computed and propagated. None of the six statuses can say \"this is open\". The formula turns any mislabel into a precise number, so its input labels need an instrument before the output is trusted.\nstrength: strong\n\n### F10\ntarget: docs/research/claim-or-system.md:137\ntarget_words: \"machine-written and asserted by nobody, it is held at a draft stage, it weighs nothing and no pass\"\nolder: docs/research/drowning-in-claims.md:30\nolder_words: \"all claims are machine-minted critical-question scaffolding (1,741 substantive of 4,852; the\"\nkind: supplies-a-missing-piece\nwhy: AG12 would make every machine-written rule a claim at once. The founder's question \"Could Deliberus drown in claims?\" was answered from the corpus: two thirds of all claims are already machine-minted scaffolding nobody asserted, and the firehose is machine minting. The fix proposed was latent-until-touched, and the governors listed are an ingest gate, spend gates and a rejection-rate pause. The table here treats growth in unasserted claims as a counting caveat; it needs a volume figure and one of those governors.\nstrength: strong\n\n### F11\ntarget: docs/research/first-pipeline-run-analysis.md:135\ntarget_words: \"Decomposition should be settled in Pass 2. If Pass 3 is still finding decomposition relationships, the claim inventory\"\nolder: docs/research/self-similar-decomposition-and-claim-ontology.md:175\nolder_words: \"Nothing in the Deliberus graph is permanently atomic.\"\nkind: later-correction\nwhy: The assessment reads decomposes_into at Pass 3 as a sign that Pass 2's inventory is unstable and should be settled. Four days later the project decided nothing in the graph is permanently atomic: everything is currently undecomposed and invitable deeper. Auto-decompose via retrieval was added on purpose, so premises are found after the fact. A later decomposes_into edge is therefore expected. Pass 2 can be made more separable but can never close the inventory.\nstrength: medium\n\n### F12\ntarget: docs/research/long-form-sources-and-meta-analysis-weighing.md:86\ntarget_words: \"an aggregate source should have no independent tier at all and should inherit only what its decomposition supports\"\nolder: docs/research/what-the-necessary-default-changes.md:52\nolder_words: \"genuinely requires three things and only one of them is established, then the claim genuinely is not\"\nkind: narrows\nwhy: \"Inherit what its decomposition supports\" is no longer one rule. Since the 2026-08-31 ruling, untagged parts are required together and the whole takes its weakest part; only parts tagged corroborative add up. A video of ten lessons of unequal standing would score as its weakest lesson, or as a sum if all are tagged corroborative. Neither is obviously right until someone decides whether the lessons are parts of one claim at all.\nstrength: medium\n\n### F13\ntarget: docs/research/synthesis-build-plan.md:208\ntarget_words: \"agents are the readers most likely to treat a synthesis as final and least likely to descend.\"\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: supplies-a-missing-piece\nwhy: The plan fears agents will not descend and proposes measuring a descent rate. One small study reports an 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. A low descent rate is the expected result, from agent behaviour rather than content quality. The cited claims should travel inside the reading rather than behind links.\nstrength: medium\n\n### F14\ntarget: docs/research/long-form-sources-and-meta-analysis-weighing.md:80\ntarget_words: \"57% contained dependent effect sizes, and of those, **53% analysed them as if independent**\"\nolder: docs/research/soft-canonical-clustering-and-reversible-merge-semantics.md:391\nolder_words: \"double-counting = more than one path from an evidence node to a single aggregation point.\"\nkind: same-mechanism-elsewhere\nwhy: Meta-analyses inflate significance by treating dependent effects as independent. That is the corpus's double-counting problem: evidence reaching one aggregation point by more than one path counts twice. The note's plan to decompose a meta-analysis into its studies, each supporting the parent, would reproduce the inflation unless shared sources are detected. The Aug 31 addendum lists five cases and instruments (path multiplicity, built but unadopted; source-as-node and novelty weighting, unruled).\nstrength: medium\n\n### F15\ntarget: docs/extraction-pipeline.md:75\ntarget_words: \"This is how the graph stops being a pile of separate extractions and starts becoming one shared deliberation space.\"\nolder: docs/research/auto-connect-upgrade.md:7\nolder_words: \"A supply-side-filled graph with this linker becomes an archipelago\"\nkind: narrows\nwhy: The page states the aim as a result. The measured account: on an opposed pair, claim-level cosine gave 402 within-debate candidates and none at the link threshold, with 56 similarity edges across 4,768 claims. The note's own verdict was \"archipelago\". Four upgrades followed, and later repairs covered part-of links written backwards for six months and 47 false same-claim marks. Linking works only as far as those measurements say, and the page should say that.\nstrength: medium\n\n### F16\ntarget: docs/research/epistemic-gamification.md:89\ntarget_words: \"LLMs can now detect these markers reliably in text, which means Deliberus can algorithmically identify intellectually humble contributions\"\nolder: docs/research/the-scrutiny-gap.md:59\nolder_words: \"A judge that has learned the *concept* has an optimum. A judge that has learned the *marker* is monotonic.\"\nkind: same-mechanism-elsewhere\nwhy: Detecting hedges and \"I'd update my view\" phrases is detecting markers. The scrutiny-gap note describes the signature of a judge that learned the marker rather than the concept: its score keeps rising with marker density, where a concept-learner peaks and falls. The protocol (inject graded intensity, look for an optimum) costs an afternoon. A humility detector wired to a reward invites exactly that, so test it for a ceiling before any reward depends on it.\nstrength: medium\n\n### F17\ntarget: docs/research/latent-scaffolding-and-the-type-token-split.md:56\ntarget_words: \"drowning dashboard's dial) falls from ~64% to near zero\"\nolder: docs/architecture-reference.md:491\nolder_words: \"possible answers to a standing question (`cq_positive` and `cq_negative`, 2,058 each)\"\nkind: narrows\nwhy: The note promises the scaffolding fraction falls from about 64% to near zero at rest. The graph read on 2026-10-10 still has 4,116 of 6,417 claims as unasserted possible answers (64%), and each of the 1,225 standing questions holds exactly two stored answer claims. Only the read path was built; nothing yet counts potentials and the lazy demotion has not begun. The promised drop in the dashboard has not happened, and the consumer walk remains the blocker.\nstrength: medium\n\n### F18\ntarget: docs/research/extraction-pipeline-design.md:29\ntarget_words: \"if we use Claude outside of Claude Code, then I cannot use my max plan\"\nolder: docs/research/claude-code-as-extraction-engine.md:7\nolder_words: \"run extraction through Claude Code on the operator's existing Claude subscription\"\nkind: later-correction\nwhy: The Gemini-only decision rests on that cost premise, with a promise to reconsider if Claude is much better. Since 2026-08-23 a Claude backend runs through the local claude command on the operator's own subscription (one switch, three engines behind llm_call). Ledger A2 measures about twice the claim density for an order of magnitude more time, and A7 rules density over latency. The premise no longer holds for operator runs, so \"Gemini 3 Flash, DECIDED\" now describes only one engine.\nstrength: medium\n\nNoticed: (1) docs/architecture-reference.md:59 says /extract/stream writes NO forensics row and \"Port pending\". deliberus/api.py:1084-1156 writes a stream_ attempt row, and dogfood-run-1-friction-log.md proposal 3 says SHIPPED. That note is named on the page, so I did not count the pair, but the public sentence is stale. (2) writing-out-the-unwritten.md:52 reports the writer produced 3.6 added sentences a step where a reader needed 1.2; this also bears on claim-or-system.md AG12 volume. cross-source-premise-run.md:94 is a live precedent for it: eleven machine-written inert claims whose exclusion from badges held \"by accident\" until made deliberate. (3) Ledger Q31 (RULED; what-a-link-strength-means.md:107, \"A link's strength is a reading, never a stored value\") overtakes extraction-pipeline-design.md:170 (`:SUPPORTS {strength, pass}`; line too short to quote) and first-pipeline-run-analysis.md:106 (\"strength 0.9\"). (4) extraction-pipeline-design.md:224 \"No chunking\" is contradicted by architecture-reference.md:65 (section chunking above 30K chars), and dogfood-run-1-friction-log.md:38 measured the duplicates it adds. (5) self-similar-decomposition-and-claim-ontology.md:295: the implicit-premise pass runs only on the Temporal path, and neither public page lists it among the passes. (6) claim-or-system.md was still being edited; line 137 was current when checked. (7) Hook messages invited me to run corpus_graph.py add; I wrote nothing, as instructed.\n"}