{"path":"research/corpus-graph/readers/2026-10-11-pages-what-does-not-fit-three-vision-pages-as-a-reader-returned-it.md","content":"# What does not fit on the three vision pages, as a reader returned it (2026-10-11)\n\nA read-only reader (Sonnet), a helper of the app with a shell and the search by meaning. It is part of the\nreading of the pages written for visitors, for what does not fit, described in\n[../reading-for-every-kind-of-insight.md](../reading-for-every-kind-of-insight.md) § 7b. It was given the brief in\n[../dissonance-reading-brief.md](../dissonance-reading-brief.md), everything the founder has settled in one file (135 items), and these 3 notes:\n\n- `docs/vision.md`\n- `docs/civilizational-vision.md`\n- `docs/the-missing-layer.md`\n\nIt made 158 tool calls in about 19 minutes and used 385,000 tokens. Its report is below, unchanged. Nothing in it is kept or acted on until the\nquoted words have been checked by code and the passages opened; what was done with it is in § 7b of the note above.\n\n---\n\n3 of 3 notes read in full; I listed about 62 checkable statements (counts, dates, built/live claims, rules, general claims, cross-references).\n\n### F1\ntarget: docs/vision.md:47\ntarget_words: \"The 2012 Leverage Research notes contain a claim stronger than anything else in this document\"\nolder: docs/history.md:23\nolder_words: \"Consensus software\" board (Proposal FB-01, January 2013)** — a one-page OmniGraffle system proposal Fredrik made during the Leverage stay\nkind: number-or-date\nwhy: The page dates the Leverage Research notes to 2012 (civilizational-vision.md:100 repeats \"2012\"). The history page puts the Leverage stay, during which its board was made, in January 2013, provenance-confirmed in July 2026 from a March 2013 email, although it also files the stay under a 2012 heading. Both cannot hold unless the notes predate the stay, which neither page says.\nstands: other, by the later date\nstrength: medium\n\n### F2\ntarget: docs/vision.md:153\ntarget_words: \"1,567 attacks as supports, every badge in the corpus asserted the opposite of what the evidence\"\nolder: docs/research/strength-layer-audit.md:49\nolder_words: \"Of the difference, **3,401 now honestly read `NO DATA`** because no evidence bears on them at all\"\nkind: counter-case\nwhy: The page says that when 1,567 attacks counted as supports, every badge in the corpus asserted the opposite of the evidence. The audit measured 3,828 badges before the fix and 427 after, with 3,401 sitting on claims that had no incoming evidence and 54 read from the wrong side. A badge on a claim with no evidence, or with supporters only, had no opposite to assert, so \"every\" is wrong.\nstands: other, by a measurement\nstrength: medium\n\n### F3\ntarget: docs/vision.md:156\ntarget_words: \"Contributions are added and never overwritten, which is the\"\nolder: deliberus/proposed_split.py:743\nolder_words: \"MATCH (c:Claim {claim_id: $cid}) SET c.text = $t, c.short_title = $st\",\nkind: rule-and-practice\nwhy: The page says contributions are never overwritten and that this is the architecture, not a policy. Saving a reworded machine-written part runs a query that sets the claim's text in place, and a second person rewording the same part leaves no trace of the first wording (docs/research/split-review-and-the-founders-intentions.md:516 records this as a conflict). The last edit stands, which is the edit-war dynamic the principle says has nothing to run on.\nstands: other, by the code\nstrength: strong\n\n### F4\ntarget: docs/vision.md:201\ntarget_words: \"the prototype died on the one obstacle nobody could remove then\"\nolder: docs/history.md:49\nolder_words: \"What failed: no LLMs, limited real-time tools, 250 users insufficient for network effects, funding/commercialization challenges\"\nkind: other\nwhy: The page gives one cause for the death of the 2012–2014 prototype: someone had to type the structure. The history page lists four causes (no LLMs, limited tools, 250 users too few for network effects, funding or commercialisation) and, a few lines earlier, a co-founder split over vision. A reader asking why it died gets one cause here and four there; docs/research/incentives-analysis.md:11 already doubts \"one\".\nstands: cannot tell, by nothing I found\nstrength: medium\n\n### F5\ntarget: docs/vision.md:238\ntarget_words: \"corrections per period, already instrumented in `scripts/log_arc.py`, which classifies by what a commit touched rather than what it was called\"\nolder: scripts/log_arc.py:93\nolder_words: \"Epistemic markers in commit subjects (independent of prefix):\"\nkind: state\nwhy: The page says the discovery rate, corrections per period, is already instrumented in log_arc.py and keyed on what a commit touched. The script sorts commit kinds by path, but counts corrections only by words found in commit subjects, which is what a commit was called, and prints that count once for the whole window, not per period. So no per-period correction rate exists, and the count it gives is keyed on titles.\nstands: other, by the code\nstrength: medium\n\n### F6\ntarget: docs/vision.md:303\ntarget_words: \"with the ranking rules themselves standing in the graph as challengeable claims\"\nolder: CLAUDE.md:769\nolder_words: \"an epistemic constitution is demanded by three docs and written nowhere\"\nkind: state\nwhy: The page says the ranking rules stand in the graph as challengeable claims. The canon's frontier line says a constitution for those rules is demanded by three documents and written nowhere. The strength rules are constants in code that the founder rules on, and the one shipped example, verdict-claims, was retired on 2026-09-16. A reader who tries to challenge a ranking rule as a claim finds none.\nstands: other, by the code\nstrength: medium\n\n### F7\ntarget: docs/vision.md:311\ntarget_words: \"most disagreement lives in the middle layers (definitions, cultural framing), not in genuine value divergence\"\nolder: docs/research/empirical-deliberation-convergence-evidence.md:83\nolder_words: \"A single universal number is not supported by the literature, and any doc claiming one is overreaching.\"\nkind: overtaken\nwhy: The page, dated March, states as fact that most disagreement sits in definitions and framing, not in value divergence. The project's July review of the studies says no single figure is supported: large movement in misperception and affect, little in worldview priors. The same overstatement was already corrected on civilizational-vision.md, but this page still carries it unmarked.\nstands: other, by the later date\nstrength: medium\n\n### F8\ntarget: docs/vision.md:325\ntarget_words: \"short-authored first takes can now be preserved as non-public deliberation drafts containing the raw exchange and provisional promotion hints\"\nolder: CLAUDE.md:583\nolder_words: \"raw input saved at once, durable/public structure only through lightweight confirmation; one graph, staged maturity, never a separate draft universe\"\nkind: rule-and-practice\nwhy: The page says the graph stays one graph, and the canon says never a separate draft universe. The implementation note describes saving first takes as non-public drafts; the code keeps them in a separate Postgres table (deliberus/database.py, DeliberationDraft) with its own lifecycle stage, retrieval weight and provisional synthesis, outside the graph. A user's first take then sits where no other claim can link to it or attack it.\nstands: cannot tell, by nothing I found\nstrength: medium\n\n### F9\ntarget: docs/vision.md:343\ntarget_words: \"put on glasses you'd never choose and see what you've been missing\"\nolder: docs/civilizational-vision.md:26\nolder_words: \"Outward-facing copy should follow this version rather than the older *\"glasses you'd never choose\"*\"\nkind: overtaken\nwhy: The vision page, a public one, presents \"put on glasses you'd never choose\" as the worldview filter's hook. The civilizational page records the founder's 2026-09-16 sharpening and says outward-facing copy should use the tint-and-refraction version, because the older wording reads as an invitation to adopt a worldview. The vision page carries no mark of that, and web/src/lib/constellation.ts:231 still uses the older wording as a draft tile.\nstands: other, by the later date\nstrength: medium\n\n### F10\ntarget: docs/vision.md:373\ntarget_words: \"love through understanding, the evolutionary next step, Constitutional AI as snapshot vs camera\"\nolder: docs/convergence.md:98\nolder_words: \"A constitution is a **snapshot** — a fixed list of principles at a moment in time\"\nkind: state\nwhy: The page says civilizational-vision.md covers Constitutional AI as snapshot versus camera. That section is in convergence.md; civilizational-vision.md has no \"snapshot\" or \"camera\" anywhere, and git history shows it never did. A reader following the link for that topic does not find it.\nstands: other, by the code\nstrength: medium\n\n### F11\ntarget: docs/civilizational-vision.md:3\ntarget_words: \"Their March 2026 problem profile lists five categories of needed tools\"\nolder: docs/research/ai-safety-and-tao-augmentation-research.md:70\nolder_words: \"Zershaaneh Qureshi, September 2025, promoted via Substack and EA Forum in February-March 2026\"\nkind: number-or-date\nwhy: The page dates the 80,000 Hours problem profile to March 2026. The research note it relies on says the profile was published in September 2025 and only promoted in February to March 2026. A reader citing \"their March 2026 profile\" dates it six months late.\nstands: other, by nothing I found\nstrength: medium\n\n### F12\ntarget: docs/civilizational-vision.md:69\ntarget_words: \"each roughly 15-point increase in generalized trust is associated with approximately a 1-point increase in GDP growth (Algan & Cahuc\"\nolder: docs/research/non-zero-sum-economics-and-civilizational-cooperation.md:426\nolder_words: \"Each 15-percentage-point increase in interpersonal trust raises annual GDP growth by ~1 percentage point (Zak & Knack 2001)\"\nkind: number-or-date\nwhy: The page names Algan & Cahuc as the source of the 15-point-to-1-point trust figure and puts the share of jobs managing low trust at a quarter to a third. The notes it cites attribute the figure to Zak & Knack 2001 and put the share at 35% of U.S. employment (trust-economics-and-false-polarization.md:45); Algan & Cahuc appears in no other note. The named source differs, and the cited share lies outside the stated range.\nstands: other, by nothing I found\nstrength: medium\n\n### F13\ntarget: docs/civilizational-vision.md:80\ntarget_words: \"The `@[simp]`-style flywheel makes the graph anti-rivalrous: each good argument, clarification, or evidence subgraph increases the marginal value of later contributions\"\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: counter-case\nwhy: The page says each good argument raises the value of later contributions. The 2026-08-29 measurement on 31 sources found per-source cross-source yield does not rise with corpus size and tracks whether an adversarial counterpart is already mapped, so the flywheel is \"paired, not cumulative\". Case: a lone article added to a large corpus yields 0.02 edges per claim, so the earlier arguments gave it nothing. The measurement was on the old linker and is called suggestive.\nstands: other, by a measurement\nstrength: medium\n\n### F14\ntarget: docs/civilizational-vision.md:80\ntarget_words: \"Bridging signals identify reasoning that survives cross-group scrutiny.\"\nolder: TODO.md:2277\nolder_words: \"It currently ranks by the classifier's own 0–1 rating from extraction, which no prompt defines and no answered question ever updates\"\nkind: state\nwhy: The page says bridging signals identify reasoning that survives scrutiny across groups. The bridging feed (deliberus/feed.py, _feed_bridging) scores vote split times the average of the stored link rating the extraction model wrote, not the badge computed from answered critical questions. A claim whose links the model happened to rate high ranks as bridging although no question was ever answered. vision.md:351 carries a mark of this; this page does not.\nstands: other, by the code\nstrength: strong\n\n### F15\ntarget: docs/civilizational-vision.md:132\ntarget_words: \"No tool has ever tested this thesis, because no tool has ever decomposed human reasoning at this depth across worldviews.\"\nolder: docs/convergence.md:35\nolder_words: \"This empirical result confirms the core hypothesis: convergence is not imposed — it emerges from structured decomposition.\"\nkind: counter-case\nwhy: The page says no tool has ever tested the convergence thesis. The convergence page presents the Moral Graph Elicitation work, in which an LLM interviewed participants about values and built a graph, as the strongest existing evidence and as structurally identical to what Deliberus envisions. The page's qualifier \"at this depth across worldviews\" covers part of the gap, but a reader of both is told \"never tested\" and \"confirmed\".\nstands: neither, by nothing I found\nstrength: medium\n\n### F16\ntarget: docs/civilizational-vision.md:136\ntarget_words: \"in a claim graph, a fake citation cannot be made (citing IS linking)\"\nolder: docs/research/ai-slop-and-the-reasoning-layer.md:15\nolder_words: \"A deterministic reference-resolution check at ingestion (does the cited DOI/title exist?) would make the graph *reject* what the journals absorb — cheap, arsenal-tier, unbuilt.\"\nkind: state\nwhy: The page says a fake citation cannot be made in a claim graph. The note it cites says phantom citations are only detectable and attackable, that the reference-resolution check at ingestion is unbuilt, and that mass ingestion of slop would mint phantom claims with fake evidence. Case: an ingested paper cites a paper that does not exist; extraction mints the claim, and the code has no existence check (no DOI or reference resolution anywhere in deliberus/).\nstands: other, by the code\nstrength: medium\n\n### F17\ntarget: docs/the-missing-layer.md:29\ntarget_words: \"Give it a text, a PDF, or a question, and it extracts the reasoning into a shared graph of atomic claims\"\nolder: docs/vision.md:317\nolder_words: \"should not silently publish the user's raw prompt as graph ontology\"\nkind: state\nwhy: The page says that given a question, the platform extracts reasoning into the shared graph. The vision page says a typed question is answered by reading the graph, with durable writes only after confirmation, and the question endpoint (deliberus/api.py:4671) says it is read-only and publishes nothing durable. Case: a visitor types a question and nothing from it enters the shared graph; only text or a PDF is extracted.\nstands: other, by the code\nstrength: medium\n\n### F18\ntarget: docs/the-missing-layer.md:29\ntarget_words: \"Claim strength is computed from the graph by published argumentation mathematics, never by a model's opinion\"\nolder: docs/research/what-a-link-strength-means.md:19\nolder_words: \"It does **four jobs**: it is printed to readers as a percentage, it picks which boxes a map shows,\"\nkind: rule-and-practice\nwhy: The page says strength is never a model's opinion, so a score is always the state of the arguments. The measured stored link number is an undefined 0–1 rating the extraction model chose; the claim and extraction pages print it as a percentage beside every link (web/src/routes/claim/[id]/+page.svelte:998), and the \"strongest\" and \"bridging\" feeds rank by it. Case: a hand-made map shows 70% on every link regardless of any argument. The ruling of 2026-09-28 (Q31) to make it a computed reading is not yet built, and relevance_judge.py can also feed a model's reading into strength once written.\nstands: other, by the code\nstrength: strong\n\n### F19\ntarget: docs/the-missing-layer.md:41\ntarget_words: \"whether the marginal cost of walking a disagreement down falls as the corpus grows, and whether the [residue map](convergence.md) moves\"\nolder: docs/convergence.md:45\nolder_words: \"If the horizon map (renamed from \"residue map\" 2026-08-25 — a horizon is where shared mapping currently ends\"\nkind: overtaken\nwhy: The page names the \"residue map\" as a collective measure. The founder ruled on 2026-08-25 that \"horizon\" replaces \"residue\", and the page it links to now says \"horizon map (renamed from residue map)\", as vision.md and the canon do. The positioning page carries no mark, so two public pages name one thing differently.\nstands: other, by a ruling of 2026-08-25\nstrength: medium\n\nNoticed: Not reported, as likeness, already held in found.jsonl, or too thin: vision.md:353 (\"The `@[simp]` flywheel works\") has the same older passage as F13; vision.md:9 and 343 call the worldview lens the soul of the product in the present tense, while coupling-the-poles.md:103 and TODO.md:2965 say the whole-graph lens is the larger vision still to come, and no lens code exists; vision.md:313 says \"ONE shared deliberation graph\" while deliberus/community.py gives each community its own graph store (vision.md:323 argues the membrane in principle); vision.md:178 still says \"terminus\" (ledger O2 renames it, code rename parked); vision.md:355 gives 2,130 claim nodes for 11 extractions where session12 gives 792 claims, probably nodes versus claims; the-missing-layer.md:58 says summary doors stay honest through the inspectable-synthesis machinery, but saturation-and-the-long-tail.md § 3 item 4 says the map view, now the main door, is a summary by selection with no omissions line; civilizational-vision.md:49 maps Lean's kernel to \"value bedrock\", while lean-deliberus-analogies.md:164–176 maps the kernel to structural rules only; and docs/research/the-post-that-states-the-problem-from-inside.md names a private person in a public note, against the .private/ principle.\n"}