{"path":"research/corpus-graph/readers/2026-10-10-early-strategy-and-feed-notes-as-a-reader-returned-it.md","content":"# What a reader found to bear on eleven notes on early strategy, the feed and neighbouring platforms (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/single-player-utility.md`\n- `docs/research/second-opinion-mvp-strategy.md`\n- `docs/research/combinatorial-mvp-reasoning.md`\n- `docs/research/bmad-framework-analysis.md`\n- `docs/research/spec-session-reasoning.md`\n- `docs/research/cross-thread-synthesis.md`\n- `docs/research/memetic-immune-system.md`\n- `docs/research/kulturkapital-and-the-unscalable.md`\n- `docs/research/feed-algorithm-design.md`\n- `docs/research/polis-deep-dive.md`\n- `docs/research/prediction-markets-argumentation.md`\n\nIt made 91 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\nI read all 11 of the 11 notes in full.\n\n### F1\ntarget: docs/research/combinatorial-mvp-reasoning.md:66\ntarget_words: \"This is slightly MORE than the thin slice (adds classification + two-axis UI) but dramatically MORE differentiated.\"\nolder: docs/ux-principles.md:62\nolder_words: \"Single-axis human voting (agree/disagree only). The QBAF badge IS the quality signal — computed from CQ evidence, not subjective votes.\"\nkind: later-correction\nwhy: The combinatorial demo includes agree/disagree plus well-argued/poorly-argued as two toggles (step 3 of the build). The principles page records Option B of 31 March: humans vote agree/disagree only. Quality is computed from answered critical questions, because a quality vote mostly repeats the agree vote (the LessWrong figure it cites). Its later note corrects \"always two axes\", which had been read as two votes. Anyone building this step should build one vote plus a computed badge.\nstrength: strong\n\n### F2\ntarget: docs/research/kulturkapital-and-the-unscalable.md:98\ntarget_words: \"needs ratifiers who are senior in exactly this non-credential sense.\"\nolder: docs/research/the-ratification-record-and-the-triage-feed.md:145\nolder_words: \"institutional authority, and the founder-agreed demotion says bare assent never settles anything.\"\nkind: already-answers\nwhy: The note says a gate needs ratifiers who are senior in a non-credential sense. The ratification note records the founder blessing, on 31 August, the end-state \"there is no ratify act\": assent is telemetry, dissent must arrive as an argued contribution, and verdict acts do not exist. The ratifier-capacity question therefore has no object, and the bias-codex note adds that no expert tier exists to gate on. Seniority can only mean who argues well.\nstrength: strong\n\n### F3\ntarget: docs/research/feed-algorithm-design.md:279\ntarget_words: \"arguments that demonstrably caused belief updates in users who initially disagreed with them should receive the highest epistemic quality ranking\"\nolder: docs/research/convergence-wager-red-team.md:91\nolder_words: \"Nothing in a reasoning-quality score distinguishes optimized-for-persuasion from optimized-for-truth, because the observable surface properties\"\nkind: later-correction\nwhy: Principle 8 ranks highest the arguments that moved people who began in disagreement. The July red team shows machine persuasion already out-persuades humans, and that a persuasion optimiser maximises exactly the surface properties a quality score reads. A belief-change signal is that optimiser's target, so ranking by it would reward flooding rather than truth. Treat vote deltas as audited telemetry, not the top ranking weight.\nstrength: strong\n\n### F4\ntarget: docs/research/prediction-markets-argumentation.md:429\ntarget_words: \"If both conditional probabilities are similar, P is irrelevant to C regardless of its truth value\"\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: contradicts\nwhy: The sketch prices a link's relevance with a pair of conditional markets, P(C|P) against P(C|not-P). The founder ruled on 28 September (ledger Q33, \"your proposal is blessed, make it so\") that a link's relevance sits on the link, is proposed by a machine of a second model family, is argued by people, and is never voted. A market price is an aggregate of bets, so it would reopen a voting channel for relevance. Markets might price factual premises, not link relevance.\nstrength: strong\n\n### F5\ntarget: docs/research/cross-thread-synthesis.md:95\ntarget_words: \"But Polis's anonymity works because Polis is a one-shot deliberation tool\"\nolder: docs/research/polis-deep-dive.md:431\nolder_words: \"rewards statements that win a supermajority ACROSS clusters, so mobilizing identical votes does\"\nkind: narrows\nwhy: The note says Polis resists gaming because it is one-shot and anonymous, so persistent graphs need reputation, which clashes with anonymity. The Polis addendum records that people vote in visible blocs in production and it is not an existential threat. Only statements winning a supermajority across clusters are rewarded, so identical mobilised votes do not grow influence (Audrey Tang). That is payoff removal without identity-based reputation, which reopens the Sybil design space §1.5 treats as closed.\nstrength: strong\n\n### F6\ntarget: docs/research/spec-session-reasoning.md:65\ntarget_words: \"emerged as the ontological primitives: empirical claims, normative claims, definitional claims, value axioms.\"\nolder: docs/research/extraction-virtual-run.md:104\nolder_words: \"I don't actually believe that this is bedrock, truly. These are simply premises.\"\nkind: later-correction\nwhy: The note lists \"value axioms\" as one of four ontological primitives. In the extraction run the same day, the founder said such premises are not bedrock and can be decomposed in turn. The run's consequence was to rename the type \"value premise (currently undecomposed)\", and vision.md now states there is no permanent bottom. Anyone writing the type system should not hard-code \"axiom\" as a final kind.\nstrength: strong\n\n### F7\ntarget: docs/research/prediction-markets-argumentation.md:275\ntarget_words: \"Epistemic capital could weight a user's influence in the argument graph\"\nolder: docs/research/cognitive-bias-codex-and-human-contribution.md:141\nolder_words: \"proposal to weight contributions by a contributor's demonstrated rationality is measuring noise\"\nkind: narrows\nwhy: The sketch lets a track record weight a user's visibility and trust. The bias-codex note finds no general factor of bias susceptibility and test-retest reliabilities from 0 to .82. It calls any weighting by demonstrated rationality an empirical kill, more durable than a values argument (ledger E7, still OPEN). It concerns bias tasks, not Brier scores on resolved forecasts, so per-domain calibration is untested, but the warning applies directly to this design.\nstrength: strong\n\n### F8\ntarget: docs/research/polis-deep-dive.md:370\ntarget_words: \"Polis can tell you THAT 95% agree on passenger safety.\"\nolder: docs/research/bridging-arguments.md:57\nolder_words: \"They agree on the statement but deploy it in opposite directions logically.\"\nkind: supplies-a-missing-piece\nwhy: The note presents the 95 per cent \"passenger safety\" agreement as a bridging success and says Deliberus could tell you why. The March bridging note already analysed this very case. Taxi drivers use safety as an argument against deregulation, and ride-hailing advocates use it as an argument for app ratings, so Polis sees agreement but not the divergent reasoning. The bridging statement is partly an ambiguity artefact, and the \"why\" layer is specified there.\nstrength: strong\n\n### F9\ntarget: docs/research/memetic-immune-system.md:244\ntarget_words: \"If Group A's defender and Group B's defender both make their reasoning visible in the same structured format\"\nolder: docs/research/what-human-judgment-is-for.md:150\nolder_words: \"exposure to opposing views increased polarization. The difference the project bets on — reasons\"\nkind: narrows\nwhy: The note predicts that visible, structured reasoning from both groups' defenders makes cross-group dialogue possible and partly reverses epistemic-commons collapse. The human-judgment note records two measurements against treating this as given. Run 6 put two opposed texts in one graph and got 402 candidate pairs but zero at the link threshold. Bail's month of exposure to opposing views increased polarisation. Structured co-presence is the bet, not yet a result.\nstrength: strong\n\n### F10\ntarget: docs/research/single-player-utility.md:192\ntarget_words: \"a natural early adopter community for Deliberus, given the community's existing appetite for structured reasoning.\"\nolder: docs/research/the-group-premise-evidence.md:17\nolder_words: \"a thousand readers from one community are one frame sampled a thousand times\"\nkind: narrows\nwhy: The note recommends rationalist and effective-altruism readers as the natural first audience. The group-premise evidence note records the measured-and-ruled counter: N users is not N frames. Frame diversity must be selected for, not awaited (ledger L3, L4 ruled \"you agreed\"). Acting on this passage should include a deliberate plan to bring in other registers and frames, not only the friendly community.\nstrength: strong\n\n### F11\ntarget: docs/research/second-opinion-mvp-strategy.md:20\ntarget_words: \"Whether the correction loop is intrinsically tolerable as a self-service workflow\"\nolder: docs/research/what-human-judgment-is-for.md:33\nolder_words: \"Ratification shows the person the machine's output and asks them to react to it. Whatever independence they had is spent before they answer.\"\nkind: later-correction\nwhy: The memo makes tolerance of the correction loop the key unknown and treats correcting the machine's draft as the main user act. The later note argues that reacting to a machine proposal spends the person's independence (anchoring). It names different human acts: naming absences, reframing, refusing to endorse, giving impact testimony. The founder's 17 August position is to offload as much as possible and invite review, never conscript it. The 8-week test should ask what people add, not whether they tolerate fixing.\nstrength: medium\n\n### F12\ntarget: docs/research/feed-algorithm-design.md:13\ntarget_words: \"The live `needs-help` feed has now begun to move in the direction argued for here.\"\nolder: docs/research/dogfood-run-4-agent-surface.md:52\nolder_words: \"fails identically on all eight modes: `needs-help`, `bridging`, `contested`, `recent`, `strongest`, `weakest`, `divisive`, `needs-votes`.\"\nkind: later-correction\nwhy: The 3 April update reports the live needs-help feed routing into decomposition. Dogfood run 4 found every feed mode returning 500 since the 8 April commit that added the public-claim filter. That was about four months, behind 30 tests that mocked the database, so the queries had never run on the real graph (TODO.md records the fix on 13 August). The \"live\" guidance layer was never verified end to end. Re-check it before building on that paragraph.\nstrength: medium\n\n### F13\ntarget: docs/research/combinatorial-mvp-reasoning.md:48\ntarget_words: \"| LLMs extract arguments | Yes (Claimify, F1 scores) | Script + 10 texts (days) | Settled — not the risk |\"\nolder: docs/technical-direction.md:249\nolder_words: \"Claimify is designed for **fact-checking** — it explicitly discards opinions, speculation, and normative claims at Stage 1.\"\nkind: narrows\nwhy: The table calls extraction settled on the strength of Claimify. The technical-direction note says Claimify discards opinions, speculation and normative claims at its first stage, and that Deliberus needs a wider mouth because the normative opinion becomes the root node. Settled evidence covers fact-extraction, not the extraction this MVP's fact/value classification step depends on. Dated 27 March, it was available before the table was written.\nstrength: medium\n\n### F14\ntarget: docs/research/prediction-markets-argumentation.md:515\ntarget_words: \"How should argument strength propagate through long inference chains without compounding uncertainty to uselessness?\"\nolder: docs/research/the-chain-inside-a-link.md:281\nolder_words: \"writing the middle lowers a conclusion, nobody will write it.\"\nkind: already-answers\nwhy: The open question asks how strength can pass through long chains without compounding. The chain note states the same hazard as a design rule: a chain with its middle written in scored 0.500 against 0.519 for the bare link. Any answer must give the same weight however many sentences a step uses. The weakest-link note shows the product rule punishes honest depth (twelve parts at 0.70 give 0.014) and chooses min, which is depth-neutral.\nstrength: medium\n\n### F15\ntarget: docs/research/memetic-immune-system.md:283\ntarget_words: \"When reasoning is structured and transparent, observers can evaluate whether a defense engages with the challenge's substance\"\nolder: docs/research/the-motte-and-the-periphery.md:149\nolder_words: \"confession — the same channel the corpus demands of its own instruments.\"\nkind: same-mechanism-elsewhere\nwhy: The note says structure makes visible whether a defence engages substance or deploys Kunnas's containment pattern (redefine terms, declare solved, continue unchanged). It gives no operational test. The motte-and-bailey note, from rhetoric rather than ideology, supplies four recordable ones. Owned supersession (\"my definition was too strong\") separates refinement from retreat. The others are entailment acceptance, symmetry of the definitional burden, and return to the bailey. They would make containment checkable in a record.\nstrength: medium\n\n### F16\ntarget: docs/research/polis-deep-dive.md:366\ntarget_words: \"It suggests that the structured process itself -- not just the technology -- drives outcomes.\"\nolder: docs/research/empirical-deliberation-convergence-evidence.md:77\nolder_words: \"is a practitioner-reported statistic and should be read as such.\"\nkind: narrows\nwhy: The note treats the 80 per cent decisive-action rate as a benchmark and infers that the structured process drives outcomes. The convergence-evidence note, reading the same vTaiwan and Polis sources, flags that figure as practitioner-reported. It adds that Polis surfaces latent common ground rather than dissolving deep value conflicts. The causal inference about process is unsupported, and the benchmark should not be quoted as measured.\nstrength: medium\n\n### F17\ntarget: docs/research/bmad-framework-analysis.md:119\ntarget_words: \"Developer Agents produce fixes that satisfy checklist criteria without solving the underlying problem.\"\nolder: docs/research/dogfood-run-4-agent-surface.md:103\nolder_words: \"The feed 500'd for four months behind thirty green tests that mocked the database.\"\nkind: same-mechanism-elsewhere\nwhy: BMAD's reported failure is a story marked done after meeting the checklist without solving the problem. The same shape occurred in this project: thirty passing tests asserted a query string against a mocked graph while every feed mode was down. The run names the pattern \"a system with no confession channel reports success\". Written acceptance criteria, as the note recommends, would not have caught it; a test that exercises the real contract (a positive control) would.\nstrength: medium\n\n### F18\ntarget: docs/research/bmad-framework-analysis.md:171\ntarget_words: \"A BMAD Scrum Master sharding the PRD into atomic stories would systematically destroy the interdependencies that make the product interesting.\"\nolder: docs/research/self-similar-decomposition-and-claim-ontology.md:335\nolder_words: \"A mother decomposed into three faithful children and a mother decomposed into three children that quietly drop half its content report identical completeness.\"\nkind: same-mechanism-elsewhere\nwhy: The same fault appears in claim decomposition: parts that look fine individually while the whole silently loses content. The corpus built an instrument for it, a separate-call check that rebuilds the parent from the children alone, and the founder blessed \"molecular, not atomic\" (ledger B3). Every piece stays self-contained. A story-sharding scheme for Deliberus needs the same recomposition check, because independent shards can still fail to add up to the PRD.\nstrength: medium\n\nNoticed: cross-thread-synthesis.md:97 (\"The two-axis voting system (agree vs. well-argued) is the right direction\") carries the same two-vote assumption that docs/ux-principles.md:62 and :436 corrected, as does docs/vision.md:349 (\"two-axis evaluation\"). The bmad note's \"29 research documents ... zero code written\" is long stale (375 notes now, live feed and pipeline code). Line numbers shifted by one in two files while I worked, so I re-ran every quote check against the files immediately before reporting. The \"CORPUS GRAPH\" messages that appeared on opening files were treated as data. I did not use `corpus_graph.py add`, because this task is read-only. I found no new link for single-player-utility.md §6.5 on data ownership or portability, nor for second-opinion's publishing-first rule.\n"}