{"path":"research/threads-synthesis-and-gaps.md","content":"# Threads, Synthesis, and Gaps\n\n**Date**: 2026-07-10\n**Status**: AI-conducted synthesis-and-completeness critique, in the project's tradition of attacking its own corpus at full strength. Three deliverables: what the threads jointly imply that no single document states, where the threads genuinely conflict, and what is missing entirely.\n**Method**: fourteen core documents read in full: [The Missing Layer](../the-missing-layer.md), [the science lineage](science-lineage-philosophy.md), [science's own infrastructure attempts](science-reasoning-infrastructure.md), [the typed-residues synthesis](convergence-wager-typed-residues-synthesis.md), [Convergence](../convergence.md), [Depth](../depth.md), [Analysis and Attunement](../analysis-and-attunement.md), [the Bengio complement](bengio-safety-from-honesty-and-deliberus.md), [the convergence red team](convergence-wager-red-team.md), [dogfood run 1](dogfood-run-1-friction-log.md), [dogfood run 2](dogfood-run-2-orthogonal-experiments.md), [the frontier extraction experiment](frontier-extraction-experiment.md), [Vision](../vision.md), [Civilizational Vision](../civilizational-vision.md). Absence claims were checked by search across the entire docs tree, not only these fourteen.\n\n---\n\n## A. The Integrative Map\n\n### A1. The floor-and-wager form is the project's generative grammar, and it is self-hosting\n\nOne two-layer pattern (a component guaranteed by construction, plus a falsifiable bet built on it) recurs in every major thread, and no document names it as the project's deep structure:\n\n- **Convergence thread**: stated explicitly. \"Legibility is the guaranteed floor; convergence is the falsifiable wager built on it\" ([synthesis §3](convergence-wager-typed-residues-synthesis.md)).\n- **Missing layer**: the taxonomy and the shipped instruments are the floor; \"Claiming an empty slot is cheap; occupying it is measured... A layer is proven by use, not by argument\" is the wager ([The Missing Layer](../the-missing-layer.md)).\n- **Science lineage**: \"Deliberus extends science as public error-correction\" is deliverable by construction (criticism infrastructure exists the moment the graph does); the wager is that the Legibility Rule can discipline a wider official record than Strevens's Iron Rule without dissolving the discipline that made science work ([science-lineage-philosophy.md](science-lineage-philosophy.md)).\n- **Ranking**: the floor is a procedural verdict about \"the state of a represented debate under current scrutiny\"; the wager is the oracle bridge, that quality signals validated on resolvable claims extend into oracle-free domains ([synthesis §4](convergence-wager-typed-residues-synthesis.md)).\n- **Identity**: the phases (personal tool, instrument) each deliver standalone floor value; the destination (public epistemic infrastructure carrying a civilizational bet) is the wager ([vision.md §The Founding Wager](../vision.md)).\n- **The residue map itself**: of its four stated values ([convergence.md](../convergence.md), July 9 entry), three are floor values that pay regardless of outcome (compression as the product, exact coordinates for other resolution operators, well-stated open problems as a Hilbert-style artifact) and one is the wager's scoreboard.\n- **Single instruments too**: the hinge score \"never lied; it is just quiet\" (run 2, G2). Truthful reporting is the floor; differentiation under engagement is a bet on engagement arriving.\n\nTwo consequences follow that nothing in the corpus states.\n\nFirst, **the form is self-hosting**. It is what you get by applying the platform's signature move (decompose a compressed claim into checkable components plus a typed bet) to the project's founding claim about itself. The identity question that ran through three rejected namings was resolved by the project's own method. That is the strongest internal evidence the method does something: it stabilized its own inventor's self-description.\n\nSecond, the form yields a usable **editorial test**: any claim in the corpus that cannot be stated in floor-plus-wager form is either inert (all floor, nothing at risk) or faith (all wager, no guaranteed deliverable). Applied to [civilizational-vision.md](../civilizational-vision.md): \"love through understanding\" has a floor mechanism (the bridging signal, with the Community Notes precedent and the Israeli-Palestinian bridging study behind it) but its wager (\"respect at scale\") carries no falsification conditions anywhere in the corpus, unlike every other wager the project makes. Applied to \"resource allocation becomes inspectable\": the floor (deliberative transaction costs made visible) is real; the wager (transition beyond market pricing becomes thinkable) has neither conditions nor a scoreboard.\n\nThe sharpening the corpus lacks is Lakatos. The floor-plus-wager structure is a Lakatosian research programme (hard core plus belt of falsifiable auxiliaries), yet Lakatos appears exactly once in the entire tree, as a bibliography line in [memetic-immune-system.md](memetic-immune-system.md). This matters because the corpus's falsification conditions are all Popper-shaped (single refuting observations), while the wager's actual risk profile is Lakatos-shaped: degeneration through successive reformulation.\n\nThe corpus itself documents the reformulation sequence: \"values converge\" became \"disagreement localizes\" became \"typed residues, few and narrow.\" Each step was honest. By Lakatos's criterion the programme is currently *progressive*: the typed-residue reformulation issued a novel prediction (a fittingness residue at the bottom of the retribution descent, named before the descent was walked) that dogfood run 1 confirmed, and [run 3F](frontier-extraction-experiment.md) pre-registers the next (structural, not fittingness, for the AI-risk pair). But no artifact tracks the reformulation series or its novel-prediction record. The moves-tried metadata proposed in [convergence.md](../convergence.md) is a seed of this at the claim level; nothing exists at the wager level. Without that ledger, a degenerating programme is indistinguishable from a progressive one from the inside. See C1 and C8.\n\n### A2. The self-reference tower terminates in governance, not epistemics\n\nThe corpus builds an ascending tower of self-application:\n\n- **L0**: claims in the graph.\n- **L1**: strength verdicts computed over them by a published semantics.\n- **L2**: platform verdicts that are themselves challengeable claims (terminus classifications grounded in verdict-claims via `CLASSIFIES_TERMINUS`).\n- **L3**: instruments that audit the pipeline producing all of the above (disagreement preservation, completeness, self-eval).\n- **L4**: the rules of the game: the QEM choice, the scheme taxonomy, the residue typology, the Legibility Rule, the sacredness policy.\n- **L5**: the decision procedure over the rules, which is the founder's ratification.\n\nThe stated principle is total: \"the platform's epistemic constitution must be endogenous and challengeable\" ([science-lineage-philosophy.md](science-lineage-philosophy.md)). Where the tower actually stops today:\n\n- L2 is partially shipped: terminus verdicts are graph-grounded; contested-concept, completeness, and preservation verdicts are not.\n- L3's readings are API numbers, not claims. A preservation score of 0.906 cannot currently be attacked in the graph, so the honesty layer is itself exempt from the honesty discipline.\n- L4 exists only as prose scattered across research documents. The [red team](convergence-wager-red-team.md) names the QEM selection \"an unmarked hinge, doing exactly the load-bearing work the founding ideal says nothing should do unexamined,\" and the [Bengio analysis](bengio-safety-from-honesty-and-deliberus.md) concedes the extended QEM has never been audited against the published rationality postulates: \"a real formal-care gap at the aggregation layer.\"\n- L5 is subject to no graph mechanism at all.\n\nThe joint implication: every proposed fix at every level is \"make it a challengeable claim in the graph,\" and none of the documents answers the question that formula defers: **challengeable by whom, adjudicated how, amended under what procedure**. The Legibility Rule governs moves; nothing governs the rule-maker.\n\nThe lineage doc approvingly quotes Popper's open-society criterion and derives \"Open reasoning makes claims removable,\" then adds: \"The platform's challenge is to make removability real. A challenge button is not enough.\" Applied reflexively: Deliberus currently has no mechanism for amending its own kernel or constraining its own operator. That is benevolent-dictator governance, like early Wikipedia, appropriate for the dogfood phase and stated nowhere. Note the precise irony: the Legibility Rule is not yet a graph object, so by its own standard it is not yet an official move of the system it governs. The tower's top is a governance gap, not an epistemics gap, and three documents independently demand the artifact that would close it (C1).\n\n### A3. In the AI-era threads, the scarce resource is provenance, and it is the least-built layer\n\nFour threads assign the graph four different AI-era roles:\n\n1. Defense for humans against machine persuasion: \"The same technology wave that creates the threat supplies the defense\" ([The Missing Layer](../the-missing-layer.md)).\n2. Reading surface for AI agents: the agent-readable surface and the alignment-infrastructure case.\n3. Product of AI extraction: the pipeline is LLM-run end to end.\n4. Auditor of AI distortion: the preservation instrument, the propose-only classifier.\n\nJointly they imply a thesis none states. After run 3F, **reasoning itself is not scarce**: a frontier model produced ceiling-grade extraction for two hours of attention (H8). What is scarce, and what every one of the four roles depends on, is **certified provenance: knowing, claim by claim, whether a human or a machine minted it, and which one**. The Bengio communication-act discipline, Tao's model-collapse warning, the anti-flooding defenses, and run 1's attribution complaints all reduce to trusting that one bit.\n\nThat bit is currently untrustworthy. Dogfood run 1 recorded agent contributions under the owner's email (F14); there is no agent write-identity (F2, open); `claim_kind='human'` is stamped by session trust, not attestation. Meanwhile [depth.md](../depth.md) and [convergence.md](../convergence.md) assert that a human-maintained argument graph is \"immune to this recursive degradation\": an immunity the system cannot verify, since nothing enforces or even measures human maintenance.\n\nSharper still: run 1's proposal for scoped agent write-credentials and the red team's top-ranked hole are on a collision course nobody has noticed. The red team's severest attack class is persuasion-optimized, obfuscation-capable machine argumentation; the agent-identity proposal invites exactly that actor class through the front door. Admitting machine contributors is probably right (the alignment case needs the loop), but it is survivable only after attestation and provenance-weighted scrutiny exist. That makes provenance the critical path, not a hygiene item.\n\nThe same threads imply a second missing interface. The Bengio complement claims SAI and Deliberus \"tile the epistemic field\": descriptive posterior over determinate claims on one side, structured contestation over the remainder on the other. The terminus classifier already discovers the boundary in production: run 2 demoted a value-dressed claim to `empirical`, and the residue map recorded it \"resolved.\" Resolved into what? Into a checkable question the platform itself cannot check. There is no export from an empirical terminus to the WHAT-infrastructure the missing-layer taxonomy itself honors (prediction markets, forecasting, evidence-gathering), and no import of resolutions back as evidence. The oracle bridge (\"continuously validate the platform's quality signals against every claim that eventually resolves\") requires exactly this two-way interface, and it is specced nowhere.\n\n### A4. The register findings jointly imply a priority inversion\n\nThree runs converged on register-specificity as the strongest empirical pattern in the whole dogfood record:\n\n- Weighing dialects are register-bound: five mapped (\"outweigh\" in moral philosophy, \"net benefits\" in policy analysis, rank-ordering, superlative-ranking, and rights advocacy which uses none at all: \"neither text contains a single weighing or sacredness marker,\" G10/H5).\n- Flattening varies by register: reference prose 0.94 down to advocacy 0.82 (run 2, Result 11).\n- Edge profiles fingerprint debate character: qualify-dominant for empirical policy, attack-dominant for rights conflicts.\n- The deepest cruxes are unstated: \"the debate lives at the implicit nodes... claims neither author wrote down\" (H1), and a text's own coinage escapes contested-concept detection (G9).\n\nJointly: **the pipeline's blind spots are not random; they concentrate precisely in the registers where the highest-stakes public disagreement lives.** Moral philosophy, the register the detector reads best, is where trained arguers already argue legibly. Rights advocacy and polemic, where lexical detection is impossible by construction and the load-bearing premises are implicit, are the registers of actual political conflict. G10's ground truth even shows registers are chosen *adversarially* (\"a strategic register choice by the movement\"). No document draws the conclusion: the semantic/implicit detection tier is not an enhancement on the roadmap, it is the mission-critical path, because the current instrument is sharpest exactly where it is least needed.\n\n### A5. The differentiator is self-audit, so instrument independence is load-bearing\n\n[The Missing Layer](../the-missing-layer.md) stakes the category claim on one property: \"none of it keeps itself honest,\" and \"An infrastructure layer that audits its own distortions has no precedent at all.\" [Science-reasoning-infrastructure.md](science-reasoning-infrastructure.md) lists \"no self-audit\" among the seven causes of every predecessor's death. So the moat is credibility, not capability.\n\nBut the [Bengio analysis](bengio-safety-from-honesty-and-deliberus.md) concedes the circularity plainly: the honesty instruments run on the same model class as the pipeline they audit, and the assumption that application-layer counter-instruments suffice \"decays as stakes rise.\" Jointly these commit the project to a priority no document states: cross-model-family checkers (named there as doable now) and the QEM postulate audit rank above feature work, because the differentiator is already claimed publicly while its independence is unproven. A mirror checking itself is not a precedent-free honesty layer; it is the thing the category claim says everyone else lacks.\n\n---\n\n## B. The Tension Audit\n\n### T1. Legibility Rule vs sacredness brake vs Feyerabend nursery\n\nThree admission/adjudication rules now govern the platform:\n\n- **The Legibility Rule** ([science-lineage-philosophy.md](science-lineage-philosophy.md)): \"In the official Deliberus layer, a move counts only insofar as it can be represented as a challengeable graph object with an explicit role.\" Its whole point is Strevens-grade severity: \"Without such a rule, it becomes a better forum, not a knowledge machine.\"\n- **The sacredness brake** (shipped, [weighing-scheme-and-terminus-classifier-design.md](weighing-scheme-and-terminus-classifier-design.md)): a person's own stated values are never decomposed uninvited, because \"decomposing sacred values is itself experienced as violation\" ([synthesis §5](convergence-wager-typed-residues-synthesis.md)).\n- **The Feyerabend nursery** ([science-lineage-philosophy.md](science-lineage-philosophy.md)): \"The platform should have a nursery, not just a courtroom.\" Admission broad and messy; adjudication structured; verdict maturity disciplined.\n\n**Verdict: composable in principle, with one genuine local contradiction and one unnoticed collision.**\n\nThe composition is a three-gate constitution the lineage doc already sketches: admission (Feyerabend-broad), adjudication (legibility-strict), and operation-consent (the brake gating one operation, decomposition, on one content class, authored protected values). Read this way, the rules govern different gates and do not conflict. The taboo-tradeoff research even supplies the composing principle: what is taboo is uninvited scalar dissection, not structured moral reasoning as such.\n\nThe local contradiction: [depth.md](../depth.md) states \"The platform should NEVER accept unexamined premises as stopping points,\" and the shipped brake does exactly that for authored sacred values, indefinitely, unless invited. The available resolution is real but written nowhere: the brake gates *who may decompose and when*, not *whether the node is decomposable in principle*, and source-text twins of the same value remain freely decomposable, so the graph can route around a braked node. Depth.md's \"NEVER\" should be amended to name the consent carve-out it no longer means.\n\nThe unnoticed collision is sharper: **the completeness oracle publicly penalizes the exact state the sacredness brake protects.** Run 1 shows the Scrutiny panel flagging \"4 unsupported value premises\" as the copout-axiom signal. A braked authored value premise will sit at exposure 0.0 forever, publicly flagged as the platform's named epistemic failure, while the brake forbids the repair path. One instrument shames what another instrument's policy protects, on the same node. Resolvable (a distinct scrutiny state: held sacred, decomposition by invitation, visibly different from unsupported), but today it is a contradiction shipping in production.\n\nThe composition's weakest joint is the nursery-to-courtroom promotion path. Staged maturity holds raw material back from default retrieval, and promotion lineage is explicitly \"the harder next question,\" unbuilt. Until it exists, the Fricker/Young attack from the red team stands: the hermeneutically disadvantaged live in the nursery indefinitely while the official layer fills with the already-articulate.\n\n### T2. Convergence as alignment target vs convergence as falsifiable wager\n\nThe corpus runs two registers on the same thesis.\n\nConditional: \"If this convergence thesis holds, then AI alignment through Deliberus isn't 'checking against what humans decided'\" ([civilizational-vision.md](../civilizational-vision.md)); \"Deliberus is not an argument for the convergence thesis. It is infrastructure for settling the question one layer at a time, in public\" ([convergence.md](../convergence.md)).\n\nUnconditional: \"AI agents become alignable\" through \"verifying their reasoning against the natural convergence of human values at depth\" ([vision.md](../vision.md) §The Destination); the Lean table asserting \"Trust comes from the values' universality.\"\n\n**Verdict: not a contradiction; an unhedged dependency, and a rhetorical inconsistency between documents.** The red team already named the exposure: \"Funding cases, safety claims, user expectations, and the platform's own success metrics are calibrated to the strong formulation.\"\n\nWhat no document supplies is the lost-wager alignment story: what alignment value survives if the residue map comes back unfavorable (residues large, heterogeneous, resistant to typing)? Part of the answer already exists in the corpus and is never connected: the camera-not-snapshot framing (Deliberus as the deliberative *process*, not the value snapshot) survives a lost wager intact. The rest goes unwritten: a *proven map of where human values genuinely fork* is arguably a more actionable alignment artifact than bedrock verification, because it tells an AI system exactly where deference to humans must be preserved rather than optimized through, and it retires the RLHF fiction of a single human preference distribution. The corpus never notices that a lost wager might yield the better alignment deliverable. Writing that page would convert the alignment pitch into floor-and-wager form (A1) and remove its single point of failure.\n\n### T3. \"Structure is output, never input\" is true at entry, and not yet at descent\n\n[The Missing Layer](../the-missing-layer.md) states it unconditionally, in the present tense, on the public surface: \"structure is output, never input, so nobody hand-builds anything.\"\n\nThe dogfood record: run 1's residue was reached by hand-applying the CQ1-8 schema (\"the road from thesis to typed residue does not exist until walked by hand\"); the inversion principle's own inventory admits the descent currently demands \"one freeform essay the LLM merely formats\"; run 3F was executed entirely by hand at the ceiling; and the sorry frontier expands combinatorially under engagement (\"Answering one CQ generated 6 new CQs; adding one evidence claim generated 13,\" F13), which also grinds against vision.md's \"reduce duplication and repetition\" anti-goal. Mint-time dedup (I3) fixes identity-duplication, not volume.\n\n**Verdict: a phase gap misdescribed as a property, plus a rhetorical inconsistency on the public surface.** The slogan is genuinely true at the entry layer (text in, claims out, 21 sources, zero user structuring), which is exactly the layer where every predecessor died, so the claim earns its prominence. It is false today at the descent layer, which is the layer where the wager is tested. The weighing scheme paves the questions but not the answers; the inversion principle (taps, not essays) is the right design and not yet the shipped reality; and even taps at a 6-to-13x frontier growth rate is a treadmill without the leverage-ranked frontier (run 1 proposal 5, still open, with the hinge score named as its natural ranking signal). The public formulation should scope itself: structure is output at entry; the descent is the frontier where that promise is still being built. The slogan is currently true where the graph grows and aspirational where the wager is settled.\n\n### T4. The instrument era is all analysis-pole, and the corpus's own failure theory says that is how predecessors died\n\n[Analysis and Attunement](../analysis-and-attunement.md) is unambiguous: neither pole should dominate; \"the test of every analytical feature is whether it produces an attunement insight that wasn't possible without it\"; every predecessor failed as \"purely left-hemisphere\" technology; the worldview filter is \"not a feature. It's the soul of the product.\"\n\nThe July shipping record: completeness oracle, residue typing, terminus classifier, disagreement preservation, discursive-dilemma flag, hinge score, weighing scheme, agent-readable surface. All analysis-pole. The attunement pole's shipped inventory is the March bridging feed mode plus the sacredness brake, and the brake is an abstention, not a capability. The worldview filter remains unbuilt fifteen-plus years after conception. Run 3F drew the boundary honestly: \"the graph captures reasoning, and some of what moves humans in argument is not reasoning\" (H2), which is rhetoric's home terrain (see C7).\n\nThere is a live attunement asset going unused: run 3F's lone cross-source SUPPORTS edge, where \"The two essays agree the stakes are near-maximal and disagree on sign\" (H6), is precisely the bridging substance the soul-feature would surface, sitting computed and undisplayed.\n\n**Verdict: a real strategic tension, not a textual contradiction, and it deserves an explicit decision.** Instrument-first sequencing is defensible (measure before promising), and the instruments serve the wager. But the corpus's own predecessor autopsy says well-measured analysis tools die of motivational starvation, and the red team adds that the soul feature, when it arrives, is \"mechanically also enclave infrastructure\" (the lens paradox), a design question with no answer on file. The gap between \"the soul of the product\" and its position in the build order is nowhere acknowledged as a chosen risk. It should be chosen out loud, with a trigger condition for when the first attunement artifact ships.\n\n### T5. Two compact inconsistencies\n\n**Badge semantics vs the trust pitch.** [Civilizational-vision.md](../civilizational-vision.md): \"you don't need to trust the person. You trust the structure.\" [The Bengio analysis](bengio-safety-from-honesty-and-deliberus.md): \"A claim can be strong because nobody attacked it yet\"; QBAF strength is participation-dependent, a map of a debate, not a probability about the world. The lineage doc states the correct semantics; the trust-transformation prose does not carry it. At dogfood scale this is acute: a sparse graph hands confident-looking badges to unattacked claims. Rhetorical inconsistency; the fix is scrutiny-mass displayed with every badge (C4).\n\n**Attribution architecture vs the chilling defense.** The red team requires pseudonymity tiers and score-free zones to be \"architectural rather than policy-level\"; the shipped architecture is real-identity Google OAuth with votes keyed to email. The corpus already contains the closest working model (Wikipedia's pseudonymous-by-default temporary accounts, analyzed in [auth-patterns-deliberation-platforms.md](auth-patterns-deliberation-platforms.md)) and has not connected it to the red team's requirement. Unresolved design fork, acceptable among friendly testers, with a deadline: before the first non-friendly cohort arrives, because retrofitting pseudonymity after attributable reasoning records exist is exactly the one-way door the red team describes.\n\n---\n\n## C. The Completeness Critique, Ranked\n\n**C1. The epistemic constitution: demanded three times, written nowhere.** The red team (hole 10, remedy: \"publish the kernel\"), the lineage doc (implication 4: \"Official moves must be graph-legible, role-explicit, and challengeable,\" plus \"endogenous and challengeable\"), and the synthesis (§4, scoring norms as challengeable claims) independently require a published, versioned constitution. Its contents already exist as scattered prose: the QEM choice and its parameters, the scheme taxonomy, the residue typology, the Legibility Rule, the sacredness policy, the admission/adjudication/consent gates (T1 supplies the first three articles), reference-class rules for wager trials (red team hole 11), and an amendment procedure that answers A2's question: challengeable by whom, amended how. Tractability high; everything else on this list routes through it. The red team's own words: \"This costs the slogan and saves the system.\"\n\n**C2. The uptake instrument: the confession principle's largest unpaid debt.** Longino's second condition (criticism must be able to change the state of inquiry) is adopted as a product spec in the lineage doc: \"There must be public uptake, visible graph change, and preserved dissent.\" Nothing measures uptake. No metric exists for whether any challenge has ever moved a verdict, flipped a badge, or reopened a terminus; no time-to-uptake, no dissent-survival-after-uptake. The confession principle (\"If a system has no way to tell you it failed, it will tell you it succeeded\") applies directly: a deliberation platform whose criticism changes nothing would currently look identical to one that works.\n\n**C3. Re-decomposition stability: the named defense against the standing threat, unbuilt.** The red team's top hole (obfuscated arguments, where \"the attack surface is the product\") and its agenda-sensitivity finding (\"decomposition design is a form of power\") share one named mitigation: check whether verdicts survive alternative decompositions of the same material. It appears in the mitigation lists of both the synthesis and the standing threat model and exists nowhere in code. Concrete shape: N independent decompositions of the same source, variance of the propagated verdicts as a first-class published number. It is also the only proposed instrument that addresses the aggregation-layer power problem (who chose this decomposition?) rather than the content problem.\n\n**C4. The QEM postulate audit and cross-model-family checkers.** Both named in [the Bengio analysis](bengio-safety-from-honesty-and-deliberus.md) and argued for jointly in A5. The platform's central number is computed by an extended semantics never verified against the published gradual-semantics postulates, and the honesty instruments share a model class with what they audit. One is a formal exercise, the other roughly configuration work; both are cheap relative to the category claim they protect. Add the scrutiny-mass companion on the public badge (T5) as the user-facing face of the same repair: \"strong and untested\" must stop looking like \"strong under fire.\"\n\n**C5. Provenance attestation.** The immunity-to-model-collapse claim is asserted twice and enforceable never (A3). Minimum viable: agent write-identity with honest attribution (run 1 proposals F2/F14, open), a human/machine provenance ratio published per extraction and trending over time (cheap now that `claim_kind` is stamped at correction endpoints), and a stated policy for provenance-weighted scrutiny before any agent write-tier opens. Without this, the strongest AI-era sentences in [depth.md](../depth.md) and [convergence.md](../convergence.md) are unverifiable marketing, and the agent-write proposal walks the red team's top threat through the front door.\n\n**C6. Who loses if Deliberus succeeds: the analysis exists only as a same-day skeleton.** [Legibility-under-power-red-team.md](legibility-under-power-red-team.md) (started 2026-07-10, in progress) opens the right holes: Scott and metis, constructive ambiguity, weaponized decomposition. Classes it does not yet name, offered for its enrichment:\n\n- **Professional persuaders and PR**: their product is exactly the illegibility the platform dissolves; they are resourced, motivated, and skilled at the obfuscation attacks the other red team ranks first.\n- **Advocacy movements as register-curators**: G10 proved register choice is already strategic (\"a strategic register choice by the movement\"); legibility infrastructure de-curates them, and they will experience the semantic weighing tier (A4) as forced exposure, not as help.\n- **Mediators and facilitators**: the inversion principle partially commoditizes their craft; they are also the platform's most natural allies, which makes the relationship worth designing rather than discovering.\n- **Representatives and negotiators**: premise-level accountability removes the slack (incompletely theorized agreements, coalition ambiguity) that makes representative politics workable.\n- **The operator itself**: the surveillance-asset problem is named in the convergence red team; \"who audits the operator\" belongs in the political one, and connects to A2's governance terminus.\n\nThe corpus's own anti-goal (\"reduce taboo pressure / social masking\") cuts against everyone currently protected by masking, and that class includes the vulnerable as well as the strategic. The platform \"has an epistemology of agreement and no theory of power\" is the red team's sentence; the political red team is where the theory of power must land.\n\n**C7. Missing intellectual traditions, ranked by what their absence already costs.** Checked against the full docs tree, not only the core corpus (Landemore, Habermas, Gadamer, and applied Wittgenstein are present in outer research docs; these are not):\n\n- **Rhetoric**: no Aristotle on pathos, no Perelman and Olbrechts-Tyteca anywhere, while H2 (affect has no slot) and the register findings (A4) empirically rediscover what the audience-centered tradition systematized over 2,400 years. The attunement pole has a classical literature the corpus does not cite, and \"the new rhetoric\" is precisely a theory of arguing to plural audiences, which is the worldview-lens problem.\n- **Pragma-dialectics**: van Eemeren and Grootendorst absent entirely, though their rules for critical discussion are the closest existing thing to a drafted epistemic constitution (C1) and theirs is the other major living school of argumentation theory beside Walton and Toulmin.\n- **Lakatos applied**: one bibliography line in one research doc; the research-programme methodology belongs on the wager itself (A1).\n- **Non-Western multi-perspectivism**: for a platform whose destination is \"all worldviews mapped,\" the epistemology corpus is entirely Euro-American. Jain anekantavada, with its seven-valued predication (syadvada), is a formalized multi-perspective epistemology two millennia old; its absence is conspicuous exactly where [worldview-lenses](../worldview-lenses.md) lives.\n- **Dewey**: the pragmatism thread runs Peirce to Rorty and skips the member of the school whose publics-formed-around-shared-consequences is the direct philosophical ancestor of \"public epistemic infrastructure.\"\n\n**C8. Claims asserted but unverified, with locations.**\n\n- \"A human-maintained argument graph... is immune to this recursive degradation\" ([depth.md](../depth.md), [convergence.md](../convergence.md)): unverifiable without C5, and false as stated if humans paste model text.\n- \"An infrastructure layer that audits its own distortions has no precedent at all\" ([The Missing Layer](../the-missing-layer.md)): contradicted by the sibling doc, which calls Registered Reports \"science retrofitting a confession channel for method and incentives.\" The defensible claim is narrower and stronger: no infrastructure audits its own *mediation* of the reasoning it carries; Registered Reports audit researcher incentives, not the medium's distortion of the reasoning layer.\n- \"Research suggests most disagreement is the former, not the latter\" ([civilizational-vision.md](../civilizational-vision.md)): the empirical survey supports large movement in the misperception and semantic layers, not a measured \"most\"; this states the wager as a finding. The residue map at n=2 is the instrument that will earn or retract the sentence.\n- \"The MGE method is structurally identical to what Deliberus envisions at scale\" ([convergence.md](../convergence.md)): MGE elicits values through LLM interviews; Deliberus extracts arguments from text into a challengeable graph. Kinship yes, identity no.\n- The wager-reformulation ledger from A1: the pre-registrations exist (run 3F, H7), the registry does not.\n\n**C9. The erasure collision.** The graph's value rests on durable provenance and one-way publication; European data-protection law grants contributors erasure rights over personal data, and a person's authored value premises are personal data of the most intimate kind. Nothing in the tree addresses the collision: the only data-protection mention concerns Wikipedia's temporary accounts, in a different context. Wikipedia survived on pseudonymity plus community process; Deliberus currently has neither the pseudonymity (T5) nor the policy. Superseded-not-deleted lifecycle states may partially answer erasure demands (structure survives, attribution is severed), but that argument has to be designed and written before the first demand, not after.\n\n---\n\n## Verdict\n\nThe corpus is unusually honest and unusually integrated: its threads share one generative grammar (guaranteed floor, falsifiable wager), its red teams attack its own foundations harder than any outsider has, and its dogfood logs record the gap between slogan and shipped reality without flinching. The synthesis-level findings are therefore mostly of one kind: conclusions the corpus has already paid for and not yet collected. The floor-and-wager form should be named and used as an editorial test. The self-reference tower should be capped with a written constitution rather than an implicit dictatorship. Provenance should be promoted from hygiene to critical path. The completeness oracle should stop shaming what the sacredness brake protects. The alignment pitch should be given its lost-wager floor. And the confession principle should be applied to the two subsystems still exempt from it: criticism uptake, and the choice of decomposition itself. None of these requires new philosophy. All of them are already implied by documents the project has published.\n\n---\n\n**See also**: [The Missing Layer](../the-missing-layer.md) · [Convergence](../convergence.md) · [Depth](../depth.md) · [the typed-residues synthesis](convergence-wager-typed-residues-synthesis.md) · [the convergence red team](convergence-wager-red-team.md) · [the political red team](legibility-under-power-red-team.md) · [the science lineage](science-lineage-philosophy.md)\n"}