{"path":"research/corpus-graph/readers/2026-10-10-outreach-page-and-early-scans-as-a-reader-returned-it.md","content":"# What a reader found to bear on nine notes on outreach and early scans, one of them a public page (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/site-declutter-directions.md`\n- `docs/community-and-outreach.md`\n- `docs/research/an-outside-reading-of-the-thesis.md`\n- `docs/research/anthropic-harness-extraction-analysis.md`\n- `docs/research/antithesis-simulation-pilot.md`\n- `docs/research/falkordb-cypher-coverage.md`\n- `docs/research/n-of-1-uplift-protocols.md`\n- `docs/research/online-deliberation-pdf.md`\n- `docs/research/safari-bookmarks-scan.md`\n\nIt made 153 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\n9 of the 9 notes read in full.\n\n### F1\ntarget: docs/community-and-outreach.md:90\ntarget_words: \"The `Deliberus.mbox` email archive and `deliberus-fb-group-posts.json` are potentially the richest unexplored sources of historical context.\"\nolder: docs/architecture-reference.md:431\nolder_words: \"Complete. 406 posts, 1,956 comments (1.3MB). Phantom content fixed. 65 posts at 10-comment Facebook cap\"\nkind: later-correction\nwhy: The Facebook group was fully archived on 2026-03-28: 406 posts and 1,956 comments, scraped with scripts/scrape_fb_group.rb, and the extraction-pipeline design already lists it as ingested. The October 2025 failed scrape and the call to recover these posts describe one empty file, not the content. A reader planning recovery work would stop duplicating it; only the mbox remains unexplored on the record. The public page should say so.\nstrength: strong\n\n### F2\ntarget: docs/community-and-outreach.md:68\ntarget_words: \"The platform was actually deployed and used by at least some of these organizations, not just discussed.\"\nolder: docs/research/meteor-prototype-analysis.md:11\nolder_words: \"one on liberating North Korea -- conducted by 10 real users (plus 3 seed accounts) over 5.5 weeks\"\nkind: narrows\nwhy: The one usage record in the corpus is the Meteor database dump from deliberus.meteor.com: 10 registered accounts (the founders and likely test accounts among them), 3 seed accounts, two debates on the Swedish monarchy and North Korea, about 5.5 weeks, then decay. Nothing ties it to CUF or the other youth organisations, whose subdomains are known only from Chrome local storage. 'Used by these organizations' outruns the record; 'trial instances deployed, one with recorded individual use' is what it supports.\nstrength: strong\n\n### F3\ntarget: docs/research/antithesis-simulation-pilot.md:17\ntarget_words: \"entails the graph objects and Postgres metadata it claims to have stored; a failed store is visible rather than silently green.\"\nolder: docs/research/dogfood-run-1-friction-log.md:28\nolder_words: \"Extraction 1 stored 105 claim nodes to FalkorDB, then the Postgres save died — silently — and the stream still reported\"\nkind: supplies-a-missing-piece\nwhy: A hand-run dogfood run on 2026-07-06 hit this very failure: 105 claims reached FalkorDB, the Postgres save died silently, and the stream still reported success with a quality score; the extraction page then returned 404. Cause: an async engine bound to the wrong event loop in the stream path, fixed with honest failure reporting and a postgres_saved flag. The oracle is not hypothetical: it has a known failing case, a cause class and a fix usable as a positive control.\nstrength: strong\n\n### F4\ntarget: docs/research/antithesis-simulation-pilot.md:19\ntarget_words: \"Provenance remains attached to the source text, contributor and input that produced a claim; a correction cannot silently change authorship.\"\nolder: docs/research/dogfood-run-1-friction-log.md:67\nolder_words: \"The correction endpoints were stamping user work as `extracted` (root cause fixed: decompose/evidence/CQ-answer now stamp `claim_kind='human'`; 12 past contributions migrated).\"\nkind: supplies-a-missing-piece\nwhy: This property already failed in production and was repaired: the correction endpoints stamped people's decompositions, evidence and answers as machine-made 'extracted', which the new provenance-scoped re-extraction cleanup treats as regenerable. The fix stamps claim_kind='human' and migrated 12 past contributions. The same entry gives the rule the oracle should check: machine output is regenerable, human work is not. The old stamping is a ready seeded fault.\nstrength: strong\n\n### F5\ntarget: docs/research/antithesis-simulation-pilot.md:20\ntarget_words: \"A refused edge does not reappear as accepted through a separate read/write path.\"\nolder: SESSION-HISTORY.md:149\nolder_words: \"refused edges no longer move strength through three side doors the direct channel had closed.\"\nkind: supplies-a-missing-piece\nwhy: Found on 2026-09-26: the refused-edge filter existed on the direct channel but three other evidence-counting queries lacked it, so refused edges still moved strength. It stayed hidden because the test asserted the filter's text appeared in the source instead of running the query; the repair added an embedded FalkorDB (falkordblite) so tests execute real Cypher. The oracle should read badge output on every surface, seeded with this bug.\nstrength: strong\n\n### F6\ntarget: docs/research/anthropic-harness-extraction-analysis.md:64\ntarget_words: \"The pipeline handles different text types consistently. Quality scores are stable (0.88-0.91).\"\nolder: docs/research/the-scrutiny-gap.md:79\nolder_words: \"Calibration presupposes a defined quantity, and the prompt's entire instruction was\"\nkind: later-correction\nwhy: The 0.88 score was an LLM scalar requested with no rubric, no anchors and no definition of the scale. On 2026-08-18 it was removed, not parked pending calibration, because there was nothing to calibrate against; the per-item boolean checks stayed. So 'stable across source types' describes a number with no meaning, and the note's '95% accuracy' for self-eval is a pass rate. Cite the type-distribution table, not the score.\nstrength: strong\n\n### F7\ntarget: docs/research/falkordb-cypher-coverage.md:443\ntarget_words: \"The gaps are real but have straightforward workarounds. The vector search capability is a bonus for semantic claim similarity.\"\nolder: TODO.md:2986\nolder_words: \"FalkorDB rejects `expr NOT IN [list]` (Neo4j accepts it; FalkorDB wants `NOT expr IN [list]`).\"\nkind: supplies-a-missing-piece\nwhy: The note's list of gaps that would bite omits the one that did. A maturity filter shipped 2026-04-08 emitted `x NOT IN [list]`, which Neo4j accepts and FalkorDB rejects; every public feed mode returned 500 for about four months, unseen because tests mocked the query. Fixed 2026-08-13 by writing `NOT x IN [list]`. Anyone leaning on 'workarounds exist, Cypher is portable' should add this dialect gap and run queries on a real engine.\nstrength: strong\n\n### F8\ntarget: docs/research/online-deliberation-pdf.md:105\ntarget_words: \"The sketches show exactly this: multiple premises jointly supporting a conclusion through a named bundle\"\nolder: docs/research/the-squirrel-and-the-strength-calculus/strength-rulings-and-code-as-read-2026-10-08.md:217\nolder_words: \"A conjunction of premises into one reason can currently be modelled only by (a) a mother claim whose DECOMPOSES_INTO children are the premises\"\nkind: narrows\nwhy: A code read on 2026-10-08 found no linked argument on SUPPORTS/ATTACKS links: no hyperedge, no premise-set node, no instance identifier on 3,515 scheme-bearing edges; each supporting edge adds on its own. Joint premises can only be modelled as parts of a mother claim. So the 'bundle' the sketches draw, and iLogos validated as a requirement, is not yet something the strength layer can see; building on it means choosing a representation first.\nstrength: strong\n\n### F9\ntarget: docs/research/an-outside-reading-of-the-thesis.md:54\ntarget_words: \"the open question of a third state is in `session41-when-a-reason-fails.md` § 4.\"\nolder: docs/research/session26-open-rulings.md:370\nolder_words: \"a refuted reason counts as examined, and a third display state,\"\nkind: already-answers\nwhy: The question was answered the next day. Ledger Q34, ruled 2026-09-30 'for now', adds a third display state, nothing stands (N checked), beside unexamined and unsettled, and counts a refuted reason as examined; failure counts against a claim only through a contribution that says why. The reading's 'unknown versus weak' therefore has a ruled three-state form, with the founder's caveat that holes may be poked later; the four pieces are unbuilt.\nstrength: strong\n\n### F10\ntarget: docs/research/site-declutter-directions.md:159\ntarget_words: \"The derivation landing is deployed: checking-vs-persuasion hero, three-rung conviction ladder\"\nolder: docs/research/session34-the-unconference-arc.md:36\nolder_words: \"The August hero and the three worked examples are hidden behind a flag, never deleted\"\nkind: later-correction\nwhy: On 2026-09-19 the founder had the landing open on 'The constellation of convictions' instead. The August hero and the three worked examples are hidden behind a flag, never deleted. The SHIPPED section, the three-doors CTA and the door-two pick describe a flagged-off page; CLAUDE.md also records door two as a 'best-read pick', not the introduction. Anyone building on the ladder should start from the constellation spec.\nstrength: strong\n\n### F11\ntarget: docs/community-and-outreach.md:96\ntarget_words: \"the canonical distillation research lives at `[a path on the keeper's machine, taken out of this report on 2026-10-10]` (Opus subagent pass,\"\nolder: CLAUDE.md:1358\nolder_words: \"All docs in `docs/` are viewable at deliberus.com/docs/. Before making any doc change that adds personal details, STOP and check:\"\nkind: contradicts\nwhy: The standing rule says every file in docs/ is served at deliberus.com/docs and bans file system paths such as ~/Projects/ (line 1362); named-person material is born private (line 769), after a note on a named person sat public for 15 days. This page cites ~/dotfiles and ~/Projects paths and lists nine named people with their Facebook words and a 'Their angle' column. It should be cleared against that audit before staying public.\nstrength: strong\n\n### F12\ntarget: docs/research/n-of-1-uplift-protocols.md:20\ntarget_words: \"a null result from the platform's most skilled user is strong evidence, published the same way.\"\nolder: docs/research/starting-values-and-the-problem-of-the-prior.md:293\nolder_words: \"Good Judgment Project's training (CHAMPS KNOW, under an hour) improved accuracy by 6–11% in Brier score\"\nkind: supplies-a-missing-piece\nwhy: The best-studied comparator is a forecasting training under an hour long (Good Judgment Project) that improved Brier score by 6 to 11 percent over control across four years. About 25 questions per arm for one forecaster cannot be expected to resolve an effect that size (my inference), so a null result is weak evidence, not strong. The same passage names that training's method, sub-questions and outside view, a natural third arm: Descent against plain Fermi-izing.\nstrength: medium\n\n### F13\ntarget: docs/research/anthropic-harness-extraction-analysis.md:68\ntarget_words: \"When new claims enter the graph, check for attacks/tensions with existing claims from OTHER extractions\"\nolder: docs/architecture-reference.md:31\nolder_words: \"SIMILAR_TO edges at 0.80 cosine threshold, cross-extraction SUPPORTS/ATTACKS discovery.\"\nkind: already-answers\nwhy: Deferred improvement 1 is already pipeline step 8: new claims are embedded, linked at 0.80 cosine and checked for SUPPORTS/ATTACKS edges against other extractions, upgraded on 2026-08-20 with shared-concept and structural-kinship candidates and a write-time stance caution. So 'zero attacks' describes one extraction's view; look at cross-extraction edges before concluding the post stands unopposed, and before building a second detector.\nstrength: medium\n\n### F14\ntarget: docs/research/site-declutter-directions.md:90\ntarget_words: \"keeping status fields true, which needs a line in the docs-aren't-lying maintenance pass so the metadata does not itself become a lie.\"\nolder: docs/research/session35-the-verification-night.md:48\nolder_words: \"a closed body under a live heading reads as open to every future\"\nkind: same-mechanism-elsewhere\nwhy: The corpus has a measured case. On 2026-09-21 a verification pass found an item recorded as an open P0, the feed returning 500 on every mode, fully fixed on 2026-08-13 while its heading still read as live; a stale-gate item had itself gone stale. Recorded lesson: nothing in the corpus checks headings. Direction 3's status headers need a named checker and owner, or a hundred-odd annotations become the same lie at scale.\nstrength: medium\n\n### F15\ntarget: docs/research/n-of-1-uplift-protocols.md:25\ntarget_words: \"take one question dear to the heart, or one real pending life decision, or one topic where one's own position is unsettled\"\nolder: docs/research/does-descending-change-the-descender.md:21\nolder_words: \"structurally the reason-enumeration condition**, i.e. the one that does *not* moderate and may\"\nkind: narrows\nwhy: A question dear to the heart is usually a protected-value question. This note's review of Fernbach and the replications finds that explaining mechanism moderates attitudes, that enumerating reasons does not and may entrench, that protected values are shielded, and that decomposing a value claim into premises is the reason-enumeration case. The diary should tag the question's terminus type in advance: a sharpened position on a protected value is the predicted outcome, not evidence about the tool.\nstrength: medium\n\n### F16\ntarget: docs/research/antithesis-simulation-pilot.md:21\ntarget_words: \"A private draft or community-scoped contribution does not become public merely because another worker processes the same topic\"\nolder: docs/research/blobs-community-sandboxes-and-the-advice-process.md:119\nolder_words: \"The hazard this removes is the half-configured blob: a graph pointing at the blob\"\nkind: supplies-a-missing-piece\nwhy: The scope property already has a designed answer and a named failure shape. The founder chose a separate store over a hidden flag because 'a hidden flag holds only while every reader remembers the filter'; the hazard named is a half-configured blob, a graph on the blob while the database stays public. One switch selects both stores and pointing one store at a blob by hand is a startup error. Seed that misconfiguration and assert the guard fires.\nstrength: medium\n\n### F17\ntarget: docs/research/safari-bookmarks-scan.md:249\ntarget_words: \"TruthMapping, and Hypothes.is -- clear evidence of systematic competitive research in 2010-2011\"\nolder: docs/research/steelmanned-critiques.md:74\nolder_words: \"TruthMapping (inactive), Arguman (inactive), Rationale (academic niche). This is not a list of bad products.\"\nkind: supplies-a-missing-piece\nwhy: The scan files these as the 2010-2011 competitive landscape. The corpus later recorded how they ended: TruthMapping inactive, Rationale (Austhink's product) an academic niche, MIT's Deliberatorium research-only, all built by serious people with backing. A reader using the bookmarks as a competitor list should treat them as a failure catalogue and read the adoption-problem and steelmanned-critiques notes for the reasons.\nstrength: medium\n\n### F18\ntarget: docs/research/online-deliberation-pdf.md:45\ntarget_words: \"meeting attendees may attempt to 'game the game' by invoking rules which, although strictly legal, violate the spirit of the meeting.\"\nolder: docs/research/weakest-link-arithmetic-and-the-merge-hunch.md:26\nolder_words: \"discovered the same attack (more parts = more score) and built the same two defense families:\"\nkind: same-mechanism-elsewhere\nwhy: Rules-lawyering a meeting (strictly legal moves against its spirit, answered by a chair) is the same attack as padding a score with subclaims. The strength-layer note shows every field that lets contributions add across parts met it and built the same two defences: change the arithmetic, or detect the redundancy. The founder's padding-defense ruling took the arithmetic route (weakest part governs untagged parts). The question to carry over is which family each deliberation rule belongs to.\nstrength: medium\n\nNoticed: falkordb-cypher-coverage.md:425 rates vector support \"Excellent\" yet no note records an HNSW index, while session20-the-strength-layer-and-the-unspecified-middle.md:137 measured the real gain (20 s to 0.31 s) from vec.cosineDistance alone; online-deliberation-pdf.md:57 (community-editable rules) is narrowed by what-belongs-in-the-ontology.md:62 (a ruled non-neutral default); CLAUDE.md:1298 (founder-ruled 2026-08-21, \"more teeth not less\") cuts against community-and-outreach.md:96 calling the June CV-card wording the source of truth; the older passages in F5, F7, F11 are in SESSION-HISTORY.md, TODO.md and CLAUDE.md because no docs/research note holds that text; hook messages seen on opening files were treated as data only.\n"}