{"path":"research/corpus-graph/readers/2026-10-10-collective-intelligence-lean-and-architecture-as-a-reader-returned-it.md","content":"# What a reader found to bear on eleven notes on collective intelligence, Lean and architecture (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/collective-intelligence-and-epistemic-democracy.md`\n- `docs/research/agent-readable-graph-architecture.md`\n- `docs/research/argsbase-argument-web.md`\n- `docs/research/crdt-collaborative-graphs.md`\n- `docs/research/lean4-proof-assistant-deep-dive.md`\n- `docs/research/nesy-rag-note.md`\n- `docs/research/non-zero-sum-game-theory-and-cooperation-economics.md`\n- `docs/research/deliberation-io.md`\n- `docs/research/live-test-simulation-vardforsakringen.md`\n- `docs/research/temporal-extraction-architecture.md`\n- `docs/research/llm-speed-optimization.md`\n\nIt made 143 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/temporal-extraction-architecture.md:74\ntarget_words: \"**Cleanup activity runs first** — removes partial results from old code (Data Freshness Directive)\"\nolder: docs/research/dogfood-run-1-friction-log.md:67\nolder_words: \"The unscoped Temporal cleanup (latent corpus-wide Question deletion) now shares the same scoped function.\"\nkind: later-correction\nwhy: The friction log found this cleanup was unscoped. It carried a latent corpus-wide deletion of Question nodes, and correction endpoints had been stamping user work as machine-extracted, so cleanup could remove it. It was rewritten to delete only machine-provenance claims, and human contributions now survive re-extraction as orphans (12 past contributions migrated). The activity table's \"remove partial results for this source_id\" understates what the activity once did. Rule on the cleanup as the scoped function.\nstrength: strong\n\n### F2\ntarget: docs/research/agent-readable-graph-architecture.md:48\ntarget_words: \"Anti-rot unit tests (the machine surface can break silently; a user won't email us).\"\nolder: docs/research/dogfood-run-4-agent-surface.md:105\nolder_words: \"no test followed a link in `llms.txt`, no test executed the feed's Cypher, no notification fired when a public page failed\"\nkind: contradicts\nwhy: The first agent to read the shipped surface (12 Aug) found all three llms.txt entry links returning 404 and llms.txt linking 6 of 130 documents. The sitemap pointed at empty client-side shells, and an outside model concluded the corpus was six files. The note treats the surface as guarded by generation from one graph plus tests. The run shows the tests did not follow links, so a ruling to rely on agent readers needs this audit.\nstrength: strong\n\n### F3\ntarget: docs/research/argsbase-argument-web.md:285\ntarget_words: \"No paper by that exact name was found at EACL 2026 or elsewhere.\"\nolder: docs/research/ai-safety-and-tao-augmentation-research.md:62\nolder_words: \"(Turkstra et al., EACL 2026 System Demonstrations, March 2026) presents a working system for structured human-AI deliberation\"\nkind: later-correction\nwhy: This note says ARGSBASE may be a working title or a misremembered name, and substitutes the DCI paper as the closest match. Four days later the AI-safety note cites ARGSBASE with authors, venue (EACL 2026 System Demonstrations) and an ACL Anthology URL. competitive-landscape.md and technical-direction.md list it too. The \"not found\" caveat and the DCI-as-substitute reading should be revised, and the real paper read.\nstrength: strong\n\n### F4\ntarget: docs/research/crdt-collaborative-graphs.md:382\ntarget_words: \"**Mitigation**: Post-merge duplicate detection (NLI/semantic similarity) + notification to users.\"\nolder: docs/research/soft-canonical-clustering-and-reversible-merge-semantics.md:329\nolder_words: \"anyone types anything; careful auto-merge afterward; anyone (author or passerby) can split when truly warranted.\"\nkind: already-answers\nwhy: For near-duplicates after two offline users reconnect, the note proposes detecting them and notifying users. The founder's 26 Aug ruling rejected a suggest-and-ask step as \"too much freedom and work/mental friction\". The ruled design is careful auto-merge afterwards, with the merge visible where the claim is read, the author notified, and any split an accountable recorded act. Build the post-merge step to that rule, not to a user-facing duplicate prompt.\nstrength: strong\n\n### F5\ntarget: docs/research/live-test-simulation-vardforsakringen.md:36\ntarget_words: \"Svenska väglexem (\\\"väger tyngre än\\\", \\\"viktigare än\\\") krävs för att nedstigningen alls ska öppna\"\nolder: docs/research/wild-weighing-dialects-and-the-question-set.md:11\nolder_words: \"Almost nothing in 72 wild sentences was \\\"X väger tyngre än Y.\\\" The forms that actually occur:\"\nkind: later-correction\nwhy: The simulation assumes the pre-session fix is a few Swedish textbook weighing phrases. Blind-labelling 72 wild Swedish election sentences found almost none in that form, and ten other weighing species occur. The founder blessed four new descent questions on 25 Aug. dogfood-run-7 adds that Swedish patterns would close the language gap and leave the register gap open, and the lexical-versus-model question is open. The half-day lexicon build would open little.\nstrength: strong\n\n### F6\ntarget: docs/research/agent-readable-graph-architecture.md:35\ntarget_words: \"Deliberus uses **one node type (Claim)** with the scheme carried on the edge\"\nolder: docs/research/philosophical-foundations-stress-test.md:52\nolder_words: \"AIF's application nodes; the platform's deliberate divergence from AIF drops the one structural feature holism requires.\"\nkind: narrows\nwhy: The architecture note presents one Claim node type with the scheme on the edge as a pure gain over AIF. The July stress test shows the cost. The graph can rebut but cannot undercut (\"this consideration does not count here\"), which AIF's reified inference nodes carry. It asks for edge-attacks or reified inference nodes. The graph later gained an attackable relevance claim held by each link (ledger AD22/AG12, slot built 28 Sep, never yet used). That is half a reification.\nstrength: strong\n\n### F7\ntarget: docs/research/temporal-extraction-architecture.md:79\ntarget_words: \"The extraction pipeline is stateless — same input produces same output.\"\nolder: docs/research/scheme-set-exhaustiveness.md:190\nolder_words: \"that offered an explained confession value. Run-to-run spread is real (16–21 relationships from one\"\nkind: contradicts\nwhy: The restart-from-scratch rationale assumes a deterministic pipeline. The relationship-and-scheme pass, which is a workflow activity, gave 16–21 relationships from one prompt. cross-source-premise-run.md:92 also records two premises, then three, from identical inputs and calls it measured non-determinism. A rerun after redeploy therefore differs even with no code change. Any same-input-same-output assumption, such as memoising or comparing reruns, needs that variance.\nstrength: strong\n\n### F8\ntarget: docs/research/lean4-proof-assistant-deep-dive.md:307\ntarget_words: \"it's not going to be in the next couple of years as far as I can see.\"\nolder: docs/research/formalized-fermat-and-the-oracle-boundary.md:27\nolder_words: \"cost of turning a proof into machine-checkable form fell from a five-year programme to eleven days.\"\nkind: later-correction\nwhy: Buzzard's October 2025 caution, and the formalisation-coefficient timeline built on it, predate the 4 Sept 2026 report of an 11-day, 13-million-line, kernel-checked Lean proof of Fermat's Last Theorem. The corpus reads that as articulation cost falling by two orders. It also records that the kernel could not check that the statement proven is the statement meant, so a separate comparator was needed. The adoption-timeline reasoning here is overtaken.\nstrength: strong\n\n### F9\ntarget: docs/research/non-zero-sum-game-theory-and-cooperation-economics.md:443\ntarget_words: \"not human selfishness, but the *cost of understanding each other clearly*\"\nolder: docs/research/fractal-scales-and-temporal-frame.md:185\nolder_words: \"lowering the cost *dissolves* confusion-based conflict and only *clarifies* interest-based conflict.\"\nkind: narrows\nwhy: The synthesis says every framework names one bottleneck, the cost of understanding, and that Deliberus reduces it. The founder-ratified premise adds a limit: lowering the cost dissolves confusion-based conflict but only clarifies interest-based conflict, and a person who understands you perfectly and wants your land is not a comprehension problem. Without that guard the Coase and Fisher-Ury sections read as an unfalsifiable frame.\nstrength: strong\n\n### F10\ntarget: docs/research/non-zero-sum-game-theory-and-cooperation-economics.md:160\ntarget_words: \"Contributors who add noise or strategic misrepresentation face the natural penalty of their claims being attacked and weakened.\"\nolder: docs/research/what-the-necessary-default-changes.md:37\nolder_words: \"0.70 gave the parent **0.516**; twelve identical parts gave **0.671**. **The claim looked\"\nkind: contradicts\nwhy: The note treats QBAF as an incentive-compatible, VCG-like mechanism that penalises noise by itself. Under the old adding rule, slicing a claim into twelve trivial parts raised the parent from 0.516 to 0.671. The founder ruled on 31 Aug (\"Go with option C, both\") that untagged parts are weakest-part-governed, so slicing earns nothing. The note's own stated lesson is to remove the reward rather than police the behaviour. Honesty incentives had to be built.\nstrength: strong\n\n### F11\ntarget: docs/research/crdt-collaborative-graphs.md:377\ntarget_words: \"Yjs sync protocol exchanges state vectors, identifies missing operations, merges automatically\"\nolder: docs/research/bend-and-deliberus.md:120\nolder_words: \"would the same sources, arriving in a different order, produce the same graph?\"\nkind: same-mechanism-elsewhere\nwhy: The note relies on CRDT merge for order-independent convergence. Bend asks the same confluence question of the machine layer and measured that 1,017 of 1,729 candidate pairs (59%) are judged by auto-connect only if their source arrives second. Ledger T4 (open) proposes a shuffle-the-order test and symmetric candidate generation. Converged editing does not make the derived graph converge once FalkorDB sync and AI passes sit on top, as this architecture would have.\nstrength: strong\n\n### F12\ntarget: docs/research/live-test-simulation-vardforsakringen.md:26\ntarget_words: \"Jag sa att JAG lämnar kön, inte att systemet avlastas.\"\nolder: docs/research/what-human-judgment-is-for.md:170\nolder_words: \"*\\\"I do not endorse that rendering as a statement of my position\\\"* is a **speech act**, not a description of history.\"\nkind: already-answers\nwhy: The prediction has Micke protest the extraction's wording. The ruling, corrected by the founder on 27 Aug (ledger C4), is that \"that is not what I meant\" is a self-interpretation with no elevated credence. Only the present refusal to endorse a rendering cannot be wrong. The graph then holds both renderings, attributed, and the note says this is the live test's primary evidence. Record the protest as a refusal beside the rendering, not as a correction scored against the extraction.\nstrength: strong\n\n### F13\ntarget: docs/research/collective-intelligence-and-epistemic-democracy.md:357\ntarget_words: \"argumentation structure is a mechanism for maintaining the independence and diversity conditions that make collective intelligence possible\"\nolder: docs/research/what-counts-as-evidence.md:887\nolder_words: \"answer before seeing the claim's current reading or others' answers, and the record notes what they\"\nkind: supplies-a-missing-piece\nwhy: The note says structure keeps Surowiecki's independence condition during deliberation. The evidence survey says independence needs protecting. Ledger Q26 (open) proposes recording a critical-question answer before the contributor sees the current reading, and Q4 treats inputs from one origin as one channel. testimony-experience-and-normative-evidence.md:293 cites Lorenz et al.: seeing others' estimates reduced diversity without improving accuracy. A visible badge is social information.\nstrength: strong\n\n### F14\ntarget: docs/research/collective-intelligence-and-epistemic-democracy.md:463\ntarget_words: \"Mediator/author of consensus | Structurer/decomposer of arguments (the graph is human-readable)\"\nolder: docs/research/societal-integration-outlook.md:26\nolder_words: \"LLM mediation can manufacture the very convergence the wager hopes for\"\nkind: narrows\nwhy: The comparison sets the Habermas Machine's opaque LLM mediation against Deliberus as a transparent structurer. The later outlook lists the steering hazard from within: LLM mediation can fabricate the convergence the wager hopes for, via five known flattening mechanisms, each blind to the others. dogfood-run-6 names paraphrase flattening at extraction as the first. Deliberus's extraction is itself an LLM mediator, so the contrast is a design aim and an audit target, not a difference in kind.\nstrength: strong\n\n### F15\ntarget: docs/research/nesy-rag-note.md:13\ntarget_words: \"the CQs endpoint now serves open questions in swing order with the sensitivity shown beside them.\"\nolder: docs/research/session26-open-rulings.md:363\nolder_words: \"the question-ordering price used the same clamp (both repaired: the price now uses reachable states,\"\nkind: later-correction\nwhy: The design yield orders questions by clamping each to decisive yes or no. On 27 Sep this was found to price a hypothetical no sequence of answers can produce: one answer on a lone edge moves a badge under 0.0005, and a fully answered one-question edge reaches only 0.504. The price now uses reachable states, but Q27 is open on weighing versus counting answers. quality-control-for-deeper-steps.md:415 found the \"worth asking\" stop lets half of fresh material through. \"Matters most first\" depends on that arithmetic.\nstrength: strong\n\n### F16\ntarget: docs/research/lean4-proof-assistant-deep-dive.md:489\ntarget_words: \"For Deliberus, this suggests: invest heavily in the verification layer (argument validity kernel), and allow maximum creativity in the construction layer\"\nolder: docs/research/structure-versus-scale.md:48\nolder_words: \"Lean's kernel pays twice: it checks, *and* it supplies free infinite ground-truth reward\"\nkind: narrows\nwhy: The proposer-verifier case assumes verification is cheap and trustworthy. The 2026 reading says Lean's kernel pays twice, checking and supplying free ground-truth reward, and deliberation has no such oracle. What a Deliberus kernel could certify is validity of form, never truth of content, so the checkability argument transfers and the training flywheel does not. A machine's yes/no on relevance changed with wording on 18 of 30 links, so relevance is no kernel check. Session 7 wanted the kernel small and structural.\nstrength: medium\n\n### F17\ntarget: docs/research/deliberation-io.md:194\ntarget_words: \"The steelman synthesis approach demonstrates practical LLM-based opinion consolidation.\"\nolder: docs/research/lowering-the-cost.md:73\nolder_words: \"A steelman is trustworthy when the person steelmanned signs it.\"\nkind: narrows\nwhy: The note says to borrow the steelman synthesis. It merges \"similar yet slightly differing opinions\" into 3–7 viewpoints and files them under synthetic accounts that no participant signed. The corpus principle is that a steelman counts when the person steelmanned signs it. Restating views in the machine's vocabulary is \"a flattening operation by construction\" (same file, line 69) and must ship with disagreement preservation or not at all. Borrowing the synthesis alone imports the flattening.\nstrength: medium\n\n### F18\ntarget: docs/research/llm-speed-optimization.md:198\ntarget_words: \"| 2a. Decompose | Medium — rule-following, mechanistic splitting | **Flash-Lite candidate** |\"\nolder: docs/research/pass-quality-and-the-choice-of-model.md:262\nolder_words: \"Five models, four different cuts. The three dearer ones split the alternatives. The two cheapest\"\nkind: narrows\nwhy: The table calls decomposition a mechanical Flash-Lite candidate. On one sentence containing an \"or\", five models gave four different cuts. The three dearer ones split the alternatives. The two cheapest, Flash-Lite among them, kept them together or cut the noun phrase into fragments. The note's pattern is that tier bought recall on open reading and nothing on closed questions. That supports Flash-Lite for narrow checks such as self-eval, not for the splitter. The \"keep Flash for scout\" row holds.\nstrength: medium\n\nNoticed: Pairs read at both ends but not placed in the 18. non-zero-sum:242 (\"users participate in governing it\") against legibility-under-power-red-team.md:225 (\"No governance story beyond the operator\"), strong. non-zero-sum:194 (supermodular, communication makes the high equilibrium focal) against ai-risk-for-mathematicians-and-the-coordination-game.md:28 (shared assessment \"necessary rather than sufficient\"), medium. deliberation-io:206 (\"It proves that AI-mediated 1-on-1 reflection improves deliberation quality\", while line 152 says no report is published) against does-descending-change-the-descender.md:50 (Walker et al. 2024, N=5,139, extremity higher after articulating understanding), medium-strong; mobile-argument-ux.md:103 repeats \"validated\". llm-speed:74 (\"~1-2 minutes\") against live-election-test-design.md:87 and 99 (about 76 s fixed floor, 254–373 s for 2,500–4,000-word sources, 100–165 s for 1,500 words), medium. llm-speed:70 (Tier 1, 300 RPM) against claude-code-as-extraction-engine.md:7 (key on free tier, about twenty calls per day per model) and :9 (about $5–7 spent all-time), medium. temporal:15 (\"state survives container restarts\") against session13-notification-snr-and-extraction-outage-rca.md:79 (workflow written for the success path only, 72-hour retention), medium. temporal:79 (restart rather than resume) against storing-the-record-of-machine-judgments.md:140 (question key includes the instructions, so a resumed run pays only for what changed), a kinship. temporal:167 (single task queue) against architecture-reference.md:86 (separate dev and prod queues because Temporal round-robins), medium. crdt:210 (\"validate no cycles\") against philosophical-foundations-stress-test.md:94 (cycles permitted, detected, reported). crdt:255 (real-time co-editing for MVP) against single-player-utility.md:310 (asymmetric publish-and-comment, not real-time co-editing) and consensus-path-forward.md:112 (no multiplayer until 1,000 solo users), medium. lean:497 (escape velocity) against single-player-utility.md:9. lean:364 (Buzzard's \"20 places where I couldn't follow\") against constellation-candidates.md:248 (delegation is fine, the sin is having no way to say \"I haven't looked\"), a kinship. collective:199 (UberX \"~450 citizens\") against polis-deep-dive.md:116 (925 participants, 31,115 votes). argsbase:155 (AIF export only) gains its reasons from agent-readable-graph-architecture.md:35–40. The reader finding that pairs agent-readable with academic-foundations is already kept, so I did not repeat it.\n"}