{"path":"research/convergence-wager-typed-residues-synthesis.md","content":"# Convergence Reaffirmed: Typed Residues, Rankable Reasoning, and the Fourth Identity\n\n**Date**: 2026-07-05\n**Status**: AI-written synthesis of a six-agent research fan-out, conducted during a live founder deliberation session. Synthesizes: [moral-convergence-metaethics-and-psychology.md](moral-convergence-metaethics-and-psychology.md), [ranking-reasoning-quality-without-bedrock.md](ranking-reasoning-quality-without-bedrock.md), [empirical-deliberation-convergence-evidence.md](empirical-deliberation-convergence-evidence.md), [convergence-wager-red-team.md](convergence-wager-red-team.md), [value-weighting-decomposition.md](value-weighting-decomposition.md), [convergence-thread-corpus-map.md](convergence-thread-corpus-map.md).\n**Nature**: research synthesis and proposal, per the project's research-as-inspiration-not-prescription boundary. The founder decides where to land.\n\n---\n\n## 1. The session this responds to\n\nFive positions were live from the founder in the July 5 session:\n\n1. **Convergence is still the wager, held fallibly.** Not \"legibility instead of convergence.\" The claim: most deep disagreement among people of good will is semantic confusion plus differently-interpreted feasibility resting on broadly shared care.\n2. **\"Instrument, not product\" is rejected** as a third successive shrinkage of the project's identity (after \"commercial product\" and \"personal tool\") — each naming the project by a smaller, more conventionally legible container than its actual destination.\n3. **Value weightings decompose too.** Ask *\"what exactly is being weighed against what?\"* and the weighting splits into constituent parts — each individually arguable. Weightings are mother claims, not bedrock.\n4. **\"We can resolve which views are better-informed and better-reasoned than others.\"**\n5. **Six candidate genuinely-unresolvable differences** tabled as test cases: retribution, partiality, population ethics, purity/sanctity, risk temperament, the value of the untouched.\n\nSix research agents then worked in parallel: metaethics and moral psychology of convergence; whether \"better-reasoned\" can be operationalized non-circularly; the empirical record of real deliberation; a dedicated red team; the weighting-decomposition move tested against all six cases; and a corpus map checking the five positions against the existing documentation.\n\n## 2. The central result: the wager survives — transformed\n\nThe strongest single finding, converged on independently by three of the six agents, is that the convergence wager survives scrutiny **in a transformed shape**. The transformation:\n\n> **From**: \"values converge when decomposed far enough.\"\n> **To**: \"decomposition converts vague weighting-disagreements into *typed, narrow, maximally legible residues* — and the wager is that the residue space is small, precisely stateable, still arguable in a different register, and empirically accounts for a minority of real-world disagreement.\"\n\nThe standard objection to convergence — *\"weightings survive full mutual understanding\"* — is not refuted but demoted. What survives full understanding, per the weighting-decomposition analysis, is not a scalar weight but a **typed residue**: a fittingness claim (retribution), a structural question about agent-relative reasons (partiality), an axiom choice under a proven impossibility theorem (population ethics), a three-way dispute about the residue's own type (purity), a risk-function choice within a formally characterized permissible band (risk temperament). These are far weaker survivors than \"bedrock values,\" because:\n\n- they are **no longer quantities** — nobody defends \"100×\" as such; the magnitude dissolves under decomposition every time;\n- they are **precisely stateable** — often a single sentence, sometimes a formal axiom;\n- they **remain arguable in a different register** — genealogical debunking versus fittingness vindication, axiom-surrender comparison, parity-commitment;\n- and decision science shows the original \"weights\" were **never stably in the head to begin with** (splitting bias, range insensitivity, method dependence, preference construction) — a claim-graph is a more psychologically realistic representation of a person's evaluative state than a utility function.\n\nPopulation ethics is the existence proof of the endgame: academic philosophy already ran the decomposition to completion, and the result is not an opaque standoff but a **formally mapped inconsistent polygon of precise axioms with a theorem proving one must be surrendered**. Bedrock-shaped, but maximally legible. If every deep disagreement could be brought to that state, the project's civilizational claim would already be redeemed — *whatever* people then choose.\n\nThe metaethics survey adds the frame that makes this respectable rather than optimistic: the strong convergence thesis (full convergence under idealized reasoning) is a minority expert position with distinguished defenders; the moderate thesis — most disagreement is semantic/factual/instrumental; structured diverse deliberation narrows it; convergence is strong at basic values and mid-level principles and weakest at terminal weightings — is well-supported. The strongest empirical pattern across cultures is **universal structure, variable weights** (Schwartz's pan-cultural value hierarchy; Curry's cooperative morals; Moral Machine's universal directional preferences). And the weighting-decomposition result says precisely that the \"variable weights\" part is the decomposable part.\n\nThe empirical record gives the honest residue profile: misperception and affective hostility (large slices of real polarization) move substantially under good conditions; complex feasibility disagreement moves some; worldview-prior residue moves very little (the Forecasting Research Institute adversarial collaboration being the hard bound — 80 median hours, mutual understanding achieved, cruxes identified, medians nearly unmoved). The wager, restated with this in hand: **the FRI residue is real but it is the last mile, not the whole road — and the platform's job is to type it, shrink it, and make both its size and its shape public.**\n\n## 3. Legibility is the floor, convergence is the wager\n\nThis session resolves a tension that had been live since May 2026. After the FRI negative result, an internal working direction had drifted toward \"legibility, not convergence\" as the project's honest central claim — make disagreement precise, claim nothing about dissolving it. The corpus map found something important: **the public documentation never actually made that retreat.** [convergence.md](../convergence.md) already states the load-bearing sentence:\n\n> \"Deliberus is not an argument for the convergence thesis. It is infrastructure for settling the question one layer at a time, in public. … 'We share more than divides us' and 'we share less than we hoped' are both real answers, and the graph has to be able to report either one without collapsing into the other.\"\n\nThe right relationship, confirmed from both directions by this fan-out: **legibility is the guaranteed floor; convergence is the falsifiable wager built on it.** Legibility is what the system delivers no matter which way the evidence goes — every dogfood session produces typed, localized disagreement. Convergence is the bet about what the accumulated residue map will show. Retreating to legibility-only would have made the project unfalsifiable and small; asserting convergence without the floor would have made it a faith claim. The two-layer statement is both honest and dangerous to its own hopes — which is what a real wager is.\n\nThe success state this implies, worth designing toward explicitly: *\"we agree on 47 of the 50 claims, and can name — with types — the two residues we differ on.\"* That sentence is simultaneously a convergence measurement, a legibility deliverable, and a bridging artifact.\n\n## 4. \"Better-reasoned\" is rankable — procedurally, with an oracle bridge\n\nThe founder's fourth claim survives with a precise shape. The operationalization stack exists and is principled: argument schemes with critical questions give typed failure modes; Bayesian argumentation makes fallacies graded and content-dependent; QBAF gradual semantics is axiomatically constrained, not arbitrary. What the resulting score honestly measures is **the state of a represented debate under current scrutiny** — not the world.\n\nTwo findings give the ranking claim teeth beyond proceduralism:\n\n- **The oracle bridge.** Where outcomes are checkable, reasoning quality demonstrably predicts accuracy: forecasting-skill correlates (active open-mindedness, granularity, comparison classes), text-visible rationale properties distinguishing top forecasters, debate experiments in which the side assigned the true answer is systematically more persuasive, and structured-reasoning training producing measurable accuracy gains. This licenses — defeasibly — extending trust in the same reasoning-quality signals into domains without oracles, and it suggests a buildable mechanism: **continuously validate the platform's quality signals against every claim that eventually resolves.**\n- **The circularity hole is real, bottomless — and universal.** There is no non-circular ground for any epistemic norms, including science's. The best available response is unusually native to this project: **put the scoring norms in the graph as challengeable claims.** A platform whose own evaluation rules are decomposable, attackable nodes has a structural answer to \"who decided what counts as better reasoning?\" that no black-box scoring system has.\n\nOne sharp caveat, named honestly: **myside bias is the one reasoning failure largely uncorrelated with reasoning disposition and intelligence**, and on identity-charged topics more reflection can polarize more. Individual epistemic virtue loses its grip exactly in the worldview domain. The implication is architectural, and it converges with the project's founding Mercier–Sperber premise: the **adversarial structure** — diverse contributors attacking each other's premises under shared rules — must carry the load that individual open-mindedness cannot. Solitary reasoning is biased; the graph is the group.\n\nSo the defensible form of position 4: *views can be ranked as better-informed and better-reasoned — as a defeasible, procedural verdict, continuously validated against resolvable claims, with the ranking norms themselves standing in the graph as challengeable claims.* Not: a truth oracle.\n\n## 5. The holes that matter most\n\nThe red team's ranked findings, with the synthesis view of each:\n\n1. **Obfuscated arguments / adversarial argumentation at scale.** Locally-valid, globally-deceptive argument structures defeat naive decomposition by construction, and persuasion-optimized generation makes them cheap. This is the sharpest structural threat *because decomposition is the platform's core mechanic — the attack surface is the product*. Mitigations are real and buildable (stability checks under re-decomposition, explicit *undecided* verdicts, bounded-verification patterns), but this hole is managed, never closed.\n2. **Convergence-illusion via LLM mediation** — judged the most underappreciated hole. Every LLM pass (extraction, decontextualization, paraphrase, auto-connect) can nudge two genuinely disagreeing humans toward the semantic center before the graph ever measures them; sycophancy and blandness bias then manufacture *measured* convergence that never happened. This corrupts the instrument in exactly the flattering direction. The [embeddings-tension](embeddings-tension-and-ai-slop.md) worry already in the corpus is the tip of it. **Buildable countermeasure: a disagreement-preservation metric on the pipeline** — round-trip checks that extracted claims preserve the original contestation.\n3. **Group-verdict incoherence (the discursive dilemma).** Premise-based aggregation (per-claim voting with strength propagation — which is literally the QBAF architecture) can provably contradict direct conclusion votes for some real opinion profiles. Most tractable of the five: display both readings and **flag the dilemma as a finding** (\"this community's premises point one way, its conclusion votes another\") rather than resolving it silently. What is a paradox for a voting system is a discovery for a legibility system.\n4. **Reasonable-pluralism residue.** Rawls's burdens of judgment — persistent reasonable disagreement *under favorable conditions* — is arguably the most widely endorsed position in political philosophy, and Chang's parity gives it structure. The typed-residue transformation (§2) is the project's honest answer: pluralism survives, but as named, narrow, typed residues rather than as an amorphous \"values differ\" blanket. The wager and reasonable pluralism are compatible until the residue map says otherwise; measuring which is the point.\n5. **The impact-capture dilemma.** If the scores never matter, the platform is inert; the more they matter, the stronger the incentive to game and capture them (Goodhart, elite capture, legibility-as-surveillance). This is a governance-layer problem that scales with success — the right time to design for it is before the scores matter, and the graph-native answer from §4 (scoring norms as challengeable claims, verdicts contestable down to their own rules) is the beginning of one.\n\nA sixth observation from the psychology of tradeoffs deserves its own line because it is an adoption risk, not an epistemology risk: **decomposing sacred values is itself experienced as violation** (taboo-tradeoff research; the mere-contemplation effect). The escape route is documented in the same literature: what is taboo is scalar comparison with the profane, not structured moral reasoning as such. Specification-style decomposition — \"what does honoring this value require here?\" — lands in the tragic/routine frame people not only tolerate but respect, and positions grounded in personal experience must be holdable as first-class content rather than flattened into propositions. This constrains the *voice* of the decomposition UX, not its possibility.\n\n## 6. The identity question: the fourth naming\n\nThree successive identities have now been rejected as legibility-driven shrinkages: **commercial product** (2013-era), **personal thinking tool** (the 2026 advisor consensus), **research instrument** (the June 2026 framing). Each named the project by its container in an existing category — SaaS, indie tool, grant-shaped apparatus — and each container was smaller than the destination.\n\nThe pattern suggests what the fourth naming must do: name the project by its **destination and its beneficiary**, not its container. The product/tool/instrument framings all answer *\"what is it for me?\"* The corpus's own north-star formulations — \"Lean for all knowledge and decision-making,\" a Wikipedia-scale deliberation graph navigable through any worldview lens, \"weaving our minds together\" — answer *\"what is it for us?\"* The category that fits is **public epistemic infrastructure**: a commons, in the lineage of Wikipedia and open-source proof libraries, that carries a falsifiable civilizational wager. Infrastructure is not a modest word — roads, courts, and encyclopedias are infrastructure — and it is the one category where \"nobody's business model\" is a feature rather than a gap.\n\nWithin that identity, the earlier framings survive as *phases* rather than rivals: it behaves like a personal thinking tool at first contact, like an instrument during the dogfood-and-measure era, and like infrastructure at destination. The naming decision is the founder's; this synthesis only observes that the fourth naming is the first one the corpus itself has been asserting all along.\n\n## 7. Buildable consequences\n\nFive concrete artifacts fall out of this fan-out, ordered by leverage:\n\n1. **Residue-typing pass.** When a disagreement is decomposed, classify what remains: *semantic / empirical / methodological / scope / aggregation-rule / risk-posture / parity-commitment / fittingness-posit* — and measure whether decomposition narrowed the disagreement. This converts every dogfood session into a data point on the founding question, natively. (The typology now exists, from the weighting-decomposition research.)\n2. **Disagreement-preservation metric.** Round-trip check on the extraction pipeline: do extracted claims preserve the original contestation, or did LLM mediation flatten it? Directly counters the most underappreciated hole.\n3. **Weighting-decomposition scheme.** A tradeoff-argument scheme with critical questions — \"What is the covering value?\", \"Is this weight a compressed empirical belief?\", \"Does it vary with framing or elicitation method?\", \"Is this a side-constraint wearing a weight's clothes?\" — turning §2's analysis into pipeline mechanics.\n4. **Oracle-bridge scoring.** Track whether high-scoring reasoning predicts resolution on claims that do resolve; publish the calibration. This is the empirical license for trusting quality signals in oracle-free domains.\n5. **Discursive-dilemma flagging.** Detect premise-vote/conclusion-vote divergence and surface it as a first-class finding.\n\n## 8. Falsification conditions, sharpened\n\nThe corpus already demands falsification conditions for the convergence thesis (the \"Hail Mary\" framing in [convergence.md](../convergence.md)). The typed-residue transformation sharpens what they can be:\n\n- **Convergence-favorable outcome**: across N decomposed disagreements between good-faith parties, the typed residue accounts for a small minority of the original disagreement's breadth; most breadth resolves at the semantic/empirical/instrumental layers; residues cluster into a small, recurring typology.\n- **Convergence-unfavorable outcome**: residues are large, heterogeneous, resistant to typing, or decomposition routinely *widens* disagreement (the motivated-reasoning backfire regime).\n- Either outcome is a discovery, and the platform must be able to report both — the floor guarantees that.\n\nThe strongest standing self-objection, inherited from the metaethics survey and left standing deliberately: the encouraging empirical evidence comes largely from structured, cooperative settings, while the platform will attract resistance-selected disputes; and philosophers — history's most practiced decomposers — have not converged in 2,500 years. The reply is not a counter-argument but the project itself: nobody has run the experiment with structure as a first-class artifact, at depth, cumulatively, in public. That is what makes it a wager rather than a claim — and what makes it worth building.\n\n---\n\n**See also**: [Convergence](../convergence.md) · [Depth](../depth.md) · [Bridging](../bridging.md) · the six underlying research docs linked in the header.\n"}