{"path":"research/corpus-graph/readers/2026-10-10-curiosity-bridging-and-adoption-as-a-reader-returned-it.md","content":"# What a reader found to bear on eleven notes on curiosity, bridging and adoption (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/curiosity-as-growth-fuel.md`\n- `docs/research/curiosity-as-connection-zurn-bassett.md`\n- `docs/research/engagement-gradient-priors.md`\n- `docs/research/adoption-problem.md`\n- `docs/bridging.md`\n- `docs/research/convergence-thread-corpus-map.md`\n- `docs/research/trust-economics-and-false-polarization.md`\n- `docs/research/ai-augmented-cooperation-infrastructure.md`\n- `docs/research/normative-reasoning.md`\n- `docs/research/bridging-arguments.md`\n- `docs/research/assumption-ranking.md`\n\nIt made 121 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/bridging.md:67\ntarget_words: \"Closing that gap changes the ranking on a shipped surface, which makes it a founder decision rather than a repair\"\nolder: docs/research/session26-open-rulings.md:367\nolder_words: \"applies *read, never set* (ruled for horizon kinds 2026-09-15) to links, answers the open bridging-feed question\"\nkind: already-answers\nwhy: Ledger row Q31 is RULED 2026-09-28. The founder answered \"Yes.\" to showing a link's strength as computed from its answered critical questions, and the row says this answers the open bridging-feed question. The public page still presents joining feed and badge as an undecided founder call. A reader should now treat the decision as made. Only the build is pending, so the page's \"not yet true\" sentence is true until that build lands.\nstrength: strong\n\n### F2\ntarget: docs/research/curiosity-as-growth-fuel.md:230\ntarget_words: \"Ratify, reject, or reword. Note that the appraisal finding makes this less fatalistic than it first reads\"\nolder: docs/research/lowering-the-cost.md:3\nolder_words: \"Curiosity is the willingness to pay the cost but the cost can also be substantially and successively lowered\"\nkind: already-answers\nwhy: Section 11 leaves the second term on the founding conviction to the founder: ratify, reject or reword. The next day the founder wrote his own sentence, \"Curiosity is the willingness to pay the cost\", and added that the cost itself can be lowered. The wording is therefore no longer unsupplied. Anyone deciding whether to write the term into vision.md should start from his sentence. frontier.md still lists the ratification as open \"Fork two\".\nstrength: strong\n\n### F3\ntarget: docs/research/bridging-arguments.md:421\ntarget_words: \"The entire bridging argument detection system depends on separately capturing reasoning quality and conclusion agreement\"\nolder: docs/ux-principles.md:62\nolder_words: \"Single-axis human voting (agree/disagree only). The QBAF badge IS the quality signal\"\nkind: later-correction\nwhy: The Option B decision of 31 March made human voting single-axis. Quality comes from answered critical questions, because a quality vote mostly repeats the agree vote (the LessWrong figure it cites is over 95 percent same-direction). This note still says dual voting axes are load-bearing and must be implemented (sections 3.2, 6.1, 7.4). Anyone building from it would build the design that was decided against. bridging.md carries the correction; this note does not.\nstrength: strong\n\n### F4\ntarget: docs/research/assumption-ranking.md:12\ntarget_words: \"We have NO empirical data on whether users can reliably separate\"\nolder: docs/research/unbuilt-features-research-mar31.md:17\nolder_words: \"Most votes go the same direction (>95% of the time)\"\nkind: already-answers\nwhy: This ranks two-axis voting as the weakest assumption, with no data. The feature note of 31 March reports LessWrong's years of production data: most votes go the same direction more than 95 percent of the time. ux-principles P2 then decided against a second human vote. So the assumption is partly measured, and the design it threatens was already dropped. Tier 1 item 1 should be rewritten as a measured result plus a decision.\nstrength: strong\n\n### F5\ntarget: docs/research/engagement-gradient-priors.md:30\ntarget_words: \"a reader who never contributes is a SUCCESS (the library test)\"\nolder: docs/research/does-descending-change-the-descender.md:124\nolder_words: \"a specific benefit that might be claimed for the artifact belongs only to the contributor tier\"\nkind: narrows\nwhy: The tier table says each tier must produce real value at that tier, and that readers are a success. The descender note finds the humility effect of decomposing needs the person to generate an explanation. Reading a finished decomposition can raise felt understanding without raising actual understanding. It says a reader of a decomposed graph gets none of that benefit. The library test survives for access, but \"real value at the reader tier\" must not claim that benefit.\nstrength: strong\n\n### F6\ntarget: docs/research/trust-economics-and-false-polarization.md:320\ntarget_words: \"evidence from identity signals → bypasses identity-protective cognition (Kahan: structure, not exhortation)\"\nolder: docs/research/curiosity-as-growth-fuel.md:17\nolder_words: \"Kahan's result says that lowering that cost for a motivated reasoner does not produce understanding. It produces a cheaper weapon.\"\nkind: later-correction\nwhy: The mechanism list says separating evidence from identity signals bypasses identity-protective cognition. The later curiosity note reads the same Kahan programme the other way. Capacity and structure served motivated reasoning, and only curiosity reversed it. Even that is shown only for information choice, not belief change (its section 10). The \"bypasses\" step should be marked unproven and conditional on curiosity. The sentence in section 5 that curiosity overrides identity protection is also stronger than that result.\nstrength: strong\n\n### F7\ntarget: docs/research/normative-reasoning.md:83\ntarget_words: \"This explains why political groups talk past each other — they are literally operating from different moral axioms.\"\nolder: docs/research/bridging-arguments.md:229\nolder_words: \"liberals and conservatives rely on very similar foundations — the apparent divergence in Haidt's original work was partly a methodological artifact\"\nkind: contradicts\nwhy: This note says the two camps operate from different moral axioms. The bridging-arguments note, from the same month, cites a 2022 APSR study. It finds both rely on very similar foundations once comparison type is controlled, and that Haidt's divergence was partly an artifact. The ethical-filter design (section 8) and the diagnostic use of foundations (section 2) both assume the divergence is large and real. Check them against this before building a foundation profile.\nstrength: strong\n\n### F8\ntarget: docs/research/trust-economics-and-false-polarization.md:293\ntarget_words: \"Curry and Alfano provide the anthropological evidence; Parfit provides the philosophical argument\"\nolder: docs/research/moral-convergence-metaethics-and-psychology.md:148\nolder_words: \"the pan-cultural value hierarchy and the seven universal moral rules are aggregate-level findings\"\nkind: narrows\nwhy: This note uses Curry and Alfano, plus Parfit, to support the claim that shared premises are far more numerous than surface discourse suggests. The metaethics note warns that these are aggregate findings. They may say little about whether two particular disputants share anything, since most variation lies within societies. It also says the encouraging evidence comes from cooperative settings, not from disputes selected for resistance, and that Parfit's convergence proof is widely judged to fail. The evidence shows a shared inventory more than shared premises.\nstrength: strong\n\n### F9\ntarget: docs/research/convergence-thread-corpus-map.md:70\ntarget_words: \"This is the position most likely to be misread by a future session, so the synthesis must be surgical.\"\nolder: docs/research/convergence-wager-typed-residues-synthesis.md:62\nolder_words: \"views can be ranked as better-informed and better-reasoned — as a defeasible, procedural verdict\"\nkind: supplies-a-missing-piece\nwhy: The map allows \"better-reasoned\" only as auditable structure and calls Position 4 in tension with the corpus. The synthesis written the same day, from this map and five other notes, gives the shape that survives. Ranking is a defeasible, procedural verdict, validated against claims that later resolve (the oracle bridge). The ranking norms themselves sit in the graph as challengeable claims. A synthesis pass should use that formulation, not the map's stricter guardrail alone.\nstrength: strong\n\n### F10\ntarget: docs/research/convergence-thread-corpus-map.md:113\ntarget_words: \"no direct anticipation found — these are the freshest three and have no existing corpus scaffolding\"\nolder: docs/research/value-weighting-decomposition.md:15\nolder_words: \"against six candidate genuinely-unresolvable differences via concrete decomposition sketches (§c)\"\nkind: already-answers\nwhy: The map reports that partiality, population ethics and the untouched have no corpus scaffolding. It also says no six-case battery is tabled. A note dated the same day tests the weighting-decomposition move against exactly those six cases. It finds four of the six leave an irreducibly evaluative residue, and it types the residues. Use that note, not the map, as the account of what already exists for the six cases.\nstrength: strong\n\n### F11\ntarget: docs/bridging.md:9\ntarget_words: \"validates that structured deliberation with typed epistemic acts significantly outperforms unstructured debate\"\nolder: docs/research/argsbase-argument-web.md:341\nolder_words: \"| DCI vs. single agent (overall) | **-0.60** | Single agent still wins on average quality |\"\nkind: narrows\nwhy: The public page treats the DCI paper as validation. The argument-web note tabulates the paper's own results. The +0.95 gain over debate holds on non-routine tasks only (n=40). DCI is the worst system on routine tasks (5.39), and overall a single agent still beats it. So \"validates\" holds for one task class against a debate baseline. It is not general validation of typed structure, and the paper is an LLM-agent simulation. The TODO.md \"citation hygiene\" item already proposes dropping DCI as a pillar.\nstrength: strong\n\n### F12\ntarget: docs/research/trust-economics-and-false-polarization.md:154\ntarget_words: \"Decompose claims into atomic sub-claims, evaluate each independently, and let the structured reasoning reveal shared premises\"\nolder: docs/research/does-descending-change-the-descender.md:172\nolder_words: \"then a value-claim decomposition may harden the decomposer. **That is the one way this project\"\nkind: narrows\nwhy: The note's response to identity-protective reasoning is to decompose and let shared premises appear. The descender note reads Fernbach and the replications. Decomposing a value claim into supporting premises is the reason-enumeration condition. That condition did not moderate attitudes, a large post-conflict study found reversal, and the effect needs self-generation. It warns decomposition may harden the decomposer while every internal instrument reads green. The claim that decomposition defuses identity threat holds, at best, for claims that bottom out in checkable fact.\nstrength: medium\n\n### F13\ntarget: docs/research/normative-reasoning.md:338\ntarget_words: \"**Respondents tend to move toward center** — effect lasts at least one year\"\nolder: docs/research/empirical-deliberation-convergence-evidence.md:36\nolder_words: \"The broader Deliberative Polling literature is more mixed than A1R alone suggests.\"\nkind: narrows\nwhy: The note generalises across 150+ experiments that shifts last a year and \"aren't transient\". The empirical-record note separates America in One Room, which persisted in a one-year follow-up, from the wider literature. In the wider record some changes reverse, the modal case shows more decay, and reason-based engagement persists better than policy attitudes. The durability takeaway should say that one strong case persisted and the general case is mixed.\nstrength: medium\n\n### F14\ntarget: docs/bridging.md:75\ntarget_words: \"Instead of requiring a physical deliberative poll, it surfaces the opposing group's strongest arguments\"\nolder: docs/research/does-descending-change-the-descender.md:201\nolder_words: \"is Deliberus's entire bet, and it must be stated as a **bet** rather than a rebuttal: nobody has run\"\nkind: narrows\nwhy: The page says a bridging architecture computationally approximates deliberative polling by showing the other side's best arguments. The descender note records Bail's month-long experiment. Following opposing views increased polarization among Republicans. The note says nobody has run that design with reasons in place of identity-marked conclusions. Fishkin's mechanism also includes moderation, balanced briefing and mixed groups. \"Approximates\" is a bet about reasoning-level exposure, not an established equivalence.\nstrength: medium\n\n### F15\ntarget: docs/research/assumption-ranking.md:38\ntarget_words: \"Whether correcting LLM-extracted argument maps feels rewarding or feels like grading homework is genuinely unknown.\"\nolder: docs/research/what-human-judgment-is-for.md:35\nolder_words: \"Ratification is the passive-acceptance mode by construction. It is the shape the evidence says produces no gain.\"\nkind: later-correction\nwhy: The ranking makes \"correction UX is the product\" depend on people tolerating correction of machine output. The human-judgment note argues, from complementarity evidence, that confirming or correcting machine proposals is the passive mode where human-AI teams gain nothing. The human acts worth asking for are absence, reframing, refusal and testimony, and the founder ratified this into canon. The thesis is replaced, so tolerability is no longer the load-bearing question.\nstrength: medium\n\n### F16\ntarget: docs/research/adoption-problem.md:322\ntarget_words: \"proved AI mediation is preferred over human mediation by participants\"\nolder: docs/research/habermas-machine.md:159\nolder_words: \"The HM's superior performance may be **partly driven by participants' misperceptions of algorithmic objectivity**\"\nkind: narrows\nwhy: \"Proved\" overstates. The dedicated reading records a 56 percent preference over untrained volunteer mediators, with equal gains in group agreement. Participants were not told the source of each statement. A Science commentary argues the advantage may be partly driven by trust in algorithmic objectivity. The system optimises approval, not argument validity. What the record supports is \"preferred in one comparison against untrained mediators\", which matters for the LLM hypothesis in sections 2 and 8.\nstrength: medium\n\n### F17\ntarget: docs/research/ai-augmented-cooperation-infrastructure.md:323\ntarget_words: \"| AI-mediated deliberation exceeds human mediation quality | Habermas Machine (n=2,110) | Strong (RCT) |\"\nolder: docs/research/synthesis-build-plan.md:93\nolder_words: \"found LLM facilitation shifted allocations by up to 5.5 percentage points while participants preferred the experience and consensus did not improve\"\nkind: narrows\nwhy: The \"proven\" table rates AI mediation quality as strong evidence, but its basis is endorsement in the Habermas study. The build-plan note reports a DeepMind study (N=879, real money). LLM facilitation steered allocations by up to 5.5 points, and participants preferred it although consensus did not improve. Preference diverged from outcome in the closest available experiment, and the steering was attributed to summarisation. The row should read \"preferred\", not \"exceeds quality\", and the steering risk belongs beside it.\nstrength: medium\n\n### F18\ntarget: docs/research/curiosity-as-connection-zurn-bassett.md:75\ntarget_words: \"the proportionality principle is master-curiosity; the sacred-domain ruling (analysis may go\"\nolder: docs/research/the-load-bearing-unsaid.md:47\nolder_words: \"Pointing the implicit-premise pass at a sacred commitment and rendering it as a weighable premise is the mechanised form of the taboo trade-off\"\nkind: same-mechanism-elsewhere\nwhy: This note treats master-curiosity (dissecting without proportionality) as a stance the sacred-domain ruling already answers. The unsaid note finds a concrete machine instance. The implicit-premise pass can surface a sacred commitment as a weighable premise. That is the mechanised taboo trade-off, which it calls a category error with a measured behavioural signature. So the Zurn warning describes a live pipeline risk. It cannot be treated as settled by a ruling that has no brake in that pass.\nstrength: medium\n\nNoticed: (1) I accidentally created an empty file /tmp/dummy with a malformed command. The permission system denied my rm, so it is still there for you to remove. I wrote nothing else. (2) Seen but not reported, or reported only partly:\n- Curiosity section 8.3 (no permanent bottom) against saturation-and-the-long-tail.md section 6b (novelty arrives ever more slowly, never stops).\n- Engagement-gradient section 3 (\"PASS for the substrate, to-do for affordances\") against interaction-modalities-and-the-pointing-test.md (unfalsifiable-falsifier risk).\n- Zurn section 2 (style signature from visit trails) against unbuilt-features-research-mar31.md:45 (prefer stated preferences over dwell and click proxies; let users pick the feed).\n- Zurn \"thing cosine cannot see\" is borne out by dogfood-run-6 J3: zero cross-domain pairs at 0.80; concepts carry the reuse, claims do not.\n- Assumption #4 against first-pipeline-run-analysis.md: the real blur was \"empirical\" overused for predictions, not fact versus value.\n- Adoption-problem's \"80% of vTaiwan led to action\" is flagged practitioner-reported in empirical-deliberation-convergence-evidence.md:75.\n- bridging.md:30 says the Konya participants \"shared a room\"; the record I found describes anonymous caucus dialogue, but that page names ai-augmented, so I excluded it. TODO.md:3525 already lists \"Bridging citation hygiene\" for bridging.md and bridging-arguments.md.\n"}