Discovery Principle and Moral Architecture — Develop-Leg Exchange

Status (June 2026): Active; Rounds 1, 2, and 4 complete (Round 3 external-human deferred; Round 5 reserved). Round 1 restates the Discovery Principle and Moral Architecture Riff (v1.2) as a synthesized set of 24 falsifiable claims (S1–S24), each paired with what would falsify it, its honest status carried from the riff (not upgraded), and the lineages that produced it. The load-bearing tier concentrates on the cross-scale transfer (S17), the dose-response make-or-break (S6/S7), the relabel/redundancy gate (S18), and the reflexive guardrail (S21); the §2.5 collapse-dynamics (S13) is carried but demoted to non-load-bearing on the cross-lineage read. A post-synthesis same-lineage coverage scan added S22–S24 (defensive/refinement claims) and three open questions (§1.9), flagged † as not part of the blind tri-lineage decomposition; a June 2026 T2 gap-close then grounded the dose-response conditions (digest Cluster C). Round 2 (adversarial) ran cross-lineage (§2): the S1–S24 set was handed to three independent model families (GPT, Grok, Gemini), blind to each other, under Adversarial Review Protocol §2 Options A+C+D — and all three returned a unanimous HOLD (no load-bearing claim survives as a supported positive finding; the decisive S17 cross-scale transfer and S18 relabel gate fail in all three). Round 4 (response/v2, §4) then converted the HOLD into a live test rather than a dead end: conceding the theory is a relabel, it relocates any novelty to a testable civic design hypothesis (§4.2) and recasts S17 as a test design with a pre-registered falsifier (§4.3). The develop-leg is now a live hold — a research program pending the §4.6 test-design memo — and nothing is promoted (the #25 hold governs, now on a hard evidentiary basis).

Why this exchange — and how Round 1 was produced. The discovery-principle riff reached v1.2 carrying claims the riff itself names as its most exposed — a risky transferable prediction it needs in order not to be apophenia (§2.4, §9 Q3), an authority-coupling / correlated-collapse claim that "owes a source-tiering pass and a small-scale test" (§9 Q11), and a make-or-break dose-response gate the draft leaves unspecified (§4.2, §8). Per riff §7 path 4, the develop-leg claims graduate into an exchange the same way the see leg graduated from the shared-mirror riff into Exchange #25. This Round 1 was built cross-lineage on purpose: the riff was decomposed independently by three model families — a same-lineage Opus baseline (control), GPT 5.5, and Gemini 3.1 — each handed the same neutral packet plus the develop-leg research-grounding digest, and the claim set below is the steward-approved synthesis of the union (medium grain). Building Round 1 this way is the direct fix for the same-lineage convergence problem that an earlier, rolled-back in-lineage attempt exhibited; the divergences across the three decompositions are recorded in §1.6 because they are more informative than the agreements.

What is not in scope. It inherits the riff §1.2 privacy firewall: idea layer only; the originating personal/biographical substrate is excluded. It promotes nothing — not a PRINCIPLES.md principle (riff path 1), not agent/doctrine/ (path 2), not a Phase 3 front-door brief (path 3) — all of which the riff gates behind these claims surviving (riff §7, §8). It does not resolve the riff §6 placement question; it carries that as S19 and tests it. And per the riff's own reflexive guardrail (§8, now itself a claim, S21), the exchange holds itself to the discipline it diagnoses: the claims ride at their honest status, and the adversarial round is reserved for a lineage that did not author them.

---

Dependencies

---

Round 1 — The develop-leg claim set under review (cross-lineage synthesis, June 2026)

1.1 What this round does

Round 1 restates the riff as a set of falsifiable claims, each paired with what would falsify it and its current honest status carried from the riff — not upgraded. A claim without a stated falsification condition is not a claim adversarial review can engage. This is the #25 Round-1 pattern applied to the develop leg.

The decomposition is a synthesis of three independent lineages (§1.6): an Opus same-lineage baseline (control), GPT 5.5, and Gemini 3.1. The grain is medium — it folds GPT 5.5's finer over-splits while preserving every theme any single lineage uniquely caught. The riff's own §2 discipline is preserved: the strong-but-near-definitional parts (S4, S8, S9) are carried as scaffolding so they cannot lend borrowed authority to the genuinely contestable claims.

1.2 The central question

(Adopted from the GPT 5.5 framing, merged with the artifact-vs-hold routing the riff §7 requires.)

Does moral architecture develop primarily from discovered framework-inadequacy rather than from instruction — and, if so, can civic architecture support graceful repair at collective scale without (a) manufacturing the crisis or (b) rebuilding authority-coupling at group scale — well enough to justify an independent artifact (riff §7 paths 1–4) over hold (path 5)?

1.3 The claim set (S1–S24)

Provenance key: G = Gemini 3.1, O = Opus baseline, X = GPT 5.5. † = added by the post-synthesis same-lineage coverage scan (§1.6), not from the three blind decompositions. "LB" = load-bearing (does the project's thesis fail if this claim fails?).

1.4 Which claims carry the weight — where Round 2's pressure belongs

Per the Adversarial Review Protocol §1 (identify the weakest load-bearing claims and argue against them as forcefully as the evidence allows), Round 2 should concentrate on:

• S17 (cross-scale transfer) — the decisive claim. All three lineages independently named this the central fault line (§1.6). The evidence base is strong for the individual and silent-or-reversed for the collective; if the transfer fails, the civic project does not follow.
• S6 + S7 (the dose-response make-or-break). The riff concedes (§4.2, §8) the civic implication is un-buildable until the growth-vs-harm conditions are specified; S7 sharpens this into a design gate.
• S18 (the relabel gate). If the Discovery Principle is transformative-learning theory relabeled for civics, its added leverage is presentational and S5/S19/S21 weaken with it.
• S21 (the reflexive guardrail). Now a falsifiable claim, not only a caution — and the one most exposed if a community framing (S16) is pursued.

The cross-lineage re-weighting of §2.5. The riff staked its non-apophenia partly on authority-coupling. Two of the three independent lineages treated the collapse-dynamics (S13) as peripheral — GPT 5.5 marked it non-load-bearing; Gemini dropped it — while all three kept S12 (pluralism-as-non-correlation). The synthesis therefore carries the durable §2.5 kernel (S12) as load-bearing and demotes S13 to a recast, non-load-bearing hypothesis, consistent with the riff's own §2.5 suddenness correction. Round 2 may re-elevate S13 if it can supply the matched-pluralist comparison the claim has never had.

The tri-lineage core — claims all three independents produced (S1, S2, S8, S10, S12, S14, S15, S17) — is the highest-confidence spine and the least likely to be a same-lineage artifact.

The coverage-scan addendum (S22–S24) is not load-bearing. These three were added after the synthesis by a same-lineage coverage scan (§1.6), not by the blind decompositions, so they carry a † and are excluded from the tri-lineage core. They are mostly defensive/refining: S22 (scaffold-vs-cage) is the frame's armor against the predictable "this is just dressed-up anti-authority / anti-religion" attack and directly buttresses S12/S15; S23 reframes S15's intervention pole; S24 is the design corollary of S10. Round 2 should pressure-test them as refinements, not weigh them as independent corroboration.

1.5 Framing flags (merged across the three lineages)

• The acquisition cluster (S1/S2/S3) may be one mechanism split three ways (acquisition / necessity / instruction-compatibility). GPT 5.5 flagged this against its own decomposition; the synthesis keeps them separate because S3's sequencing is the part the evidence most supports, but Round 2 may merge them.
• The dose-response gap (S6/S7) is the engineering hole — a dose-response model is assumed before any variables, thresholds, or measurement are specified.
• The taxonomy (S8) risks reading as exhaustive when it may be a useful typology rather than a complete account; people blend and move between responses.
• Coupling claims (S10/S11) may conflate identity-coupling, belonging threat, and authority dependence — later rounds should keep them separable. S11's fixed change-rate ordering is smuggled confidence (flagged by both GPT 5.5 and Gemini): presented as fact, it is a complex empirical claim needing its own grounding.
• S13–S12 need clean operational definitions for "authority," "common root," "correlation," and "totality."
• S15/S16 mix empirical prediction with normative guardrail and may need splitting in later drafting; whoever designs the exercises holds capture-prone power (the §8 / S21 risk).
• S17 is doing two jobs — it is both a substantive claim (the transfer holds) and a methodological discipline (only count it if it pays out collectively), which risks circularity.
• S19 is a placement hypothesis inside the project's architecture, not a claim about the world.

1.6 Cross-lineage synthesis method and the divergences that matter

Round 1 was produced under the Comparative Alignment Protocol logic: three lineages decomposed the same riff independently, blind to each other, on the same packet + research-grounding digest. The Opus output is labeled a same-lineage baseline / control, not a vote.

• Granularity spread (itself a finding). Gemini cut 8 claims, Opus 15, GPT 5.5 29. The riff has no canonical claim count, so claim boundaries are author-imposed; the synthesis picks a deliberate medium grain (~21, plus a 3-claim same-lineage coverage-scan addendum → 24 total). Gemini's 8 were a near-strict subset of the common core (it caught nothing the others missed); GPT 5.5 was the maximal decomposition; Opus sat in the middle plus the unique S18.
• Tri-lineage convergence on the attack surface. All three independently named cross-scale transfer and the dose-response gap as the first targets for an adversary — across different training lineages, which is the strongest signal in the exercise that these are real soft spots, not same-lineage artifacts.
• Unique catches preserved. GPT 5.5 alone surfaced S5 (collective fragility as a falsifiable civic claim), S3 (sequencing), S7 (the design gate), the explicit S13 "more-total-not-sudden", S17's evidence-asymmetry half, and S21 (guardrail-as-claim); it also wrote the strongest falsification conditions. Opus alone surfaced S18 (the relabel gate). Gemini contributed the coarse common-core view and the cleanest single-question framing.
• The §2.5 re-weighting (S13 demoted; S12 kept) is the most consequential divergence and is recorded in §1.4.
• The coverage-scan addendum (disclosed, not tri-lineage). After the three blind decompositions were synthesized, a same-lineage coverage scan checked S1–S21 against the full riff and added S22–S24 (scaffold-vs-cage, riff §2.5; machine↔therapist, riff §5.1; continuity-through-revision, riff §1.4) and the three §1.9 open questions — all already present in the riff as prose but un-lifted by the blind pass. These carry a †: they are a same-lineage completeness patch, not tri-lineage corroboration, and an independent Round 2 should treat them with the same skepticism as any single-lineage claim.

1.7 What the claims are not

• Claims that the Discovery Principle is correct. Correctness is a downstream (doctrine) question. The set asks whether the develop-leg claims survive structured pressure well enough to justify any independent artifact (riff §7 paths 1–4) over hold (path 5).
• A promotion of any kind. No principle, no doctrine heuristic, no Phase 3 brief. The #25 hold discipline binds here too.
• A resolution of the §6 placement question. S19 tests placement; it does not assert it.
• A re-litigation of the see leg or the triad. #25's M-claims and #26's T-series ride at their own status; this exchange uses them, it does not re-grade them.
• Claims about the steward's standpoint. The riff §1.2 privacy firewall is upstream; idea layer only.

1.8 Round plan

Rounds 1, 2, and 4 are complete (Round 2 cross-lineage, §2; Round 4 response/v2, §4). Round 4 was run ahead of Round 3 at steward direction — the external-human Round 3 is deferred, not skipped, and does not block. The develop-leg is now a live hold (a research program pending a bounded test, §4.6), not a parked one. Round 2 was run by independent lineages, not the authoring (Opus) lineage.

On request, a standalone adversary packet (the S-claims as assertions to be tested, reduced context, optional domain lens) can be produced as the Round 2 handoff to whichever independent lineage or reviewer runs it.

1.9 Open questions the coverage scan surfaced (carried into Round 2, not yet claims)

Raised by the same-lineage coverage scan (§1.6) as gaps not tied to a single claim, and carried so Round 2 weighs them rather than re-discovering them as gotchas. These are questions, not assertions — none is a claim with a falsification condition.

1. Measurement. S6/S7/S10/S12 are falsifiable in principle but name no measurement instrument. How would "architecture," "repair capacity," "coupling," or "dose" actually be operationalized — and does the digest Cluster C finding (Frazier et al. 2009: self-reported growth ≠ measured growth) apply to the whole develop leg, not just the trauma literature?
2. Temporal sequencing vs. the triad. Beyond placement (S19), does "develop" have an ordering relative to see/decide/act — a prerequisite that must precede them, a continuous substrate that runs under all three, or neither?
3. Community reversibility. What mechanism keeps a practice-bound community (S16) from drifting over time into a grievance-bound, founder-centralized, series-wired form (S21)? Belonging that enables repair and belonging that rebuilds authority-coupling may be the same structure at different times.

---

Round 2 — Independent cross-lineage adversarial review (June 6, 2026)

2.1 How it was run

Per the Adversarial Review Protocol Options A + C + D and the lineage-independence default: the S1–S24 claim set was handed out as a standalone adversary packet (reduced context + a moral-psychology / political-socialization / adult-learning skeptic lens) to three independent model families — GPT, Grok, and Gemini — each blind to the others and to the project's same-lineage interpretive read. Each re-derived the literature itself; none inherited the digest's tiers. This is the reserved adversarial round the Round-1 cross-lineage synthesis set up, and it is deliberately not run in the authoring (Opus) lineage. The three full outputs live in the steward's working thread; this section is the steward-approved synthesis.

2.2 Verdict — unanimous HOLD (3 / 3)

No load-bearing claim survives as a supported positive finding. The two decisive claims — S17 (cross-scale transfer) and S18 (relabel gate) — fail in all three lineages. The only claims marked "supported" are S7 (and only as a constraint the project currently fails) and S24† (supported but explicitly not novel). The #25 hold is confirmed and strengthened; nothing is promoted.

Per the steward decision accompanying this filing, the Round 2 statuses below are recorded as findings and are NOT applied back to the Round 1 LB / status columns (§1.3). Re-weighting the claim set is a v2 / Round 4 act, and Round 4 is not run — the exchange holds at Round 2 (§2.10).

2.3 Epistemic-status table (per Adversarial Review Protocol §3)

Cont. = contested · Uns. = unsupported · Under. = underspecified · Rel. = relabel-of-existing · Sup. = supported.

2.4 The decisive failures

• S17 (cross-scale transfer) — decisive and unmet. "Conceded and fatal" (Grok); "ecological fallacy" — societies have institutions, power, incentives, legitimacy, coalition formation, not a brain that experiences dissonance (Gemini); the leap introduces mechanisms that "are not just bigger individuals" (GPT). The civic claim rests on a metaphor, not a finding.
• S18 (relabel gate) — fails as written. All three map the load-bearing parts onto Mezirow (disorienting dilemma), Festinger (dissonance), Kahan (identity-protective cognition), pluralism / deliberation, and Edmondson (psychological safety). Survivable only if narrowed to an operational civic design hypothesis (§2.8).
• S6 / S7 — "dose-response" is the wrong abstraction. All three: it is conditions over magnitude (safety, identity-threat, voluntariness, facilitation) — confirming the digest Cluster C gap-close finding — and S7's precondition is currently unmet by the project's own admission.

2.5 Attacks the Round-1 self-assessment did NOT anticipate (the strong cross-lineage signal)

Convergence on S17/S18 is a weak signal (the project flagged them itself). These new, independently-landed attacks are the strong signal that the claim set has problems beyond the known ones:

• The Haidt post-hoc problem (3/3). Moral reasoning is largely post-hoc rationalization → "moral exercises" may train better rationalizers, not architecture. Gemini: the project cites Haidt but misapplies him. Threatens the whole acquisition cluster (S1/S2/S4).
• "Supply creates demand" (Gemini). Critical pedagogy / consciousness-raising reverses S3's demand-then-supply arrow — instruction can generate the break.
• The "privilege of readiness" (Gemini). S15's wait-for-readiness biases toward the status quo; social movements must manufacture breaks (civil disobedience).
• Grievance communities DO build architecture (3/3; Gemini sharpest). Unions, civil-rights, liberation movements — the practice-vs-grievance binary (S16, and the §5.3 community framing) is "ideologically loaded / false."
• S10 directional flexibility (Gemini). Fusion dictates the direction of repair (partisans revise to maintain fusion) rather than foreclosing it.
• S12 → anomie/apathy/factional-sorting (GPT + Gemini). Non-correlated authorities can produce Durkheimian anomie or cynicism, not synthesis.

2.6 Where the lineages diverged (their distinct value)

• GPT — most structurally generous: names narrowed S1/S3 + S6/S7 + S10/S16 as potentially survivable if operationalized; sharpest on the measurement question.
• Grok — most decision-forcing: minimal surviving set = S7 + S17 + S18, all currently failing; best pre-registration cross-examination questions; cited effect sizes (Kapur g ≈ 0.36–0.87 under high fidelity).
• Gemini — deepest theory: ecological fallacy, Haidt misapplication, supply-creates-demand, privilege-of-readiness, false grievance binary.

2.7 Reviewer-surfaced sources (not in the digest)

Eight sources named per the Research-Protocol §2 W8 discipline: Kahan (identity-protective cognition → S10); Bandura (social learning → counter to S2); Piaget (disequilibrium → S1/S2, strengthens the S18 relabel case); Latané (social impact → counter to S12); Durkheim (anomie → counter to S12); Rogers (client-centered therapy → S23 relabel); Kohlberg/Rest DIT intervention literature (→ S2 relabel); the moral-progress literature (institutional/economic drivers → S17). Now grounded (June 2026, after Round 4 opened a v2) with full citations and tiers as digest Cluster D (§9) — most cut against or bound the claims, landing hardest on the two Round 2 found decisive: named relabel antecedents for S18 (Piaget, Rogers, Kohlberg/Rest) and a competing structural account of collective moral change for S17 (Buchanan & Powell).

2.8 The one constructive path (convergent across all three)

The only survivable artifact is a narrowed civic design hypothesis, framed as a research agenda / adversarial risk memo / pre-registration scaffold — explicitly not a principle or doctrine. "A synthesis claim, not a discovered principle" (GPT); minimal surviving set S7/S17/S18 (Grok); Gemini's concrete version: design civic institutions to expose citizens to non-correlated authorities to force individual synthesis — testable, with the burden being operational definitions and comparative tests.

2.9 Cross-examination questions carried forward

The nine questions across the three converge on three themes (carried, not answered, since the exchange holds here):

1. Measurement — a behavioral measure of "moral architecture" that distinguishes it from vocabulary / openness / intelligence / ideological moderation / ordinary belief-updating (lands on §1.9 Q1).
2. Collective-scale falsification — a matched collective-scale case (or disconfirming prediction) that would force abandoning, not reinterpreting, the transfer.
3. Pre-registration / stop-conditions — published operational definitions of productive-vs-harmful tension and a stop condition, before any exercise is designed; plus Gemini's political challenges (the Haidt problem; the privilege of readiness).

2.10 Round status

Round 2 complete. The exchange did not stop here: at steward direction it proceeded to a Response round (§4) ahead of the (now deferred) external-human Round 3.

---

Round 4 (Response / v2) — Reframe to a testable civic design hypothesis (June 6, 2026)

Run out of order. This is the plan's Round 4 (response → v2), run before the external-human Round 3 (deferred until a human reviewer is on hand — it does not block). Steward-directed.

Its discipline: a response round is not a rescue. Its job is to convert the Round 2 HOLD into either a live test or an honest death — and the test it proposes must be able to come back NO. The trigger was a steward reframe (the "three-utilities" puzzle: K₃,₃ is impossible on a flat plane but solvable once you add a dimension — and, crucially, "no evidence yet" is not the same kind of impossible as a proof of impossibility). S17 was never proven false; it was ruled unwarranted for now. This round adds the missing dimension by designing the test that would earn or kill it.

4.1 The move — from claim to design hypothesis under test

Round 2 killed the develop-leg as a settled civic claim (S17 transfer unmet; S18 relabel). v2 stops carrying it as a claim to defend and carries it as a testable civic design hypothesis — the one path all three Round-2 lineages named as survivable (§2.8). Language fix adopted project-wide for this leg: retire "survives / fails" finality; use "untested, pending a named test."

4.2 S18-v2 — concede the theory, relocate the novelty

Concession (honest): as a psychological theory, the develop-leg is a relabel — Mezirow + Festinger + Kahan + Piaget + pluralism + Edmondson. The claim to a new law of the mind is dropped.

Where novelty could legitimately live — a civic design + transfer claim:

A civic body (institution, community, or polity) deliberately engineered for (1) routine exposure to non-correlated authorities/inputs, (2) psychological safety for dissent, and (3) consolidation after a shared conclusion breaks — will, when hit by a shock that invalidates a prior shared conclusion, repair (revise the broken belief while preserving function) at a higher rate than a matched answer-transmission body, without manufacturing the shock or recentralizing authority.

This is an engineering recipe plus a risky transfer prediction. Whether it is already stated is the real S18 test — plausibly not, because the source literatures are individual/organizational, not civic-self-correction-by-design. So: theory = relabel; design hypothesis = possibly novel, and testable.

4.3 S17-v2 — the test that could come back NO

S17 is recast from "the transfer holds" to "here is the smallest test whose result earns or kills the transfer."

1. Measure first (answers the Haidt post-hoc attack, §2.5). "Collective repair capacity" = on meeting a shock that invalidates a prior shared conclusion, does the body revise the belief while keeping function vs. deny/calcify, fracture, or wholesale-convert? Measured by behavior — policy actually revised under new evidence, dissent tolerated without exit, decision quality on novel dilemmas — not self-report, not vocabulary. If exercises improve justifications but not decisions, the hypothesis fails. (Frazier-style measurement caution, digest Cluster C.)
2. Beat the ecological fallacy (§2.5). Unit of analysis = the body's process (deliberation, dissent channels, decision outputs), not the average member's attitude. Measure the institution, not a pile of individuals.
3. Matched comparison. Two matched bodies — one answer-transmission, one reasoning-practice (non-correlated inputs + safety + consolidation) — same naturally-occurring shock — repair vs. fracture.
4. Pre-registered falsifier. If reasoning-practice bodies do not out-repair controls — or fracture/polarize more (the group-polarization risk realized, digest Cluster A / Sunstein) — or "win" only by manufacturing crises or leaning on a central authority — the transfer is dead. Stated up front.
5. Ethics (S7). Naturally-occurring shocks only (no manufactured crises); voluntary, opt-in exercises; stop conditions on harm.
6. Start with the smallest cylinder — which is already spinning. The project's own adversarial-review method is a micro-instance of this hypothesis: non-correlated authorities (different model lineages) + psychological safety (challenge-not-applause, no promotion pressure) + consolidation (synthesis rounds). This exchange is a data point — three independent lineages forced a HOLD that a single lineage rubber-stamped in the rolled-back attempt. That is the develop-leg mechanism operating on the project itself, and it is the first concrete test rung. Honest caveat: n = 1, it tests a project's beliefs not a society's, and the project chose the adversarial method partly believing the mechanism — so it is suggestive, not confirmatory; it is the first rung, not proof.

4.4 The other Round-2 attacks → v2 refinements (real concessions)

4.5 Changelog (v1 → v2)

4.6 Routing recommendation

• Reclassify the develop-leg: candidate principle/claim → testable civic design hypothesis (research program).
• Hold all promotion (principle / doctrine / Phase-3) pending a bounded test result. The #25 hold still governs — but it is now a live hold (parked pending a named experiment), not a dead one.
• Next artifact — drafted (June 2026): the S17 test-design memo — measurement instrument (collective repair capacity, behavioral; ecological-fallacy guard), matched comparison, the pre-registered falsifier set (F1–F6), ethics/stop conditions, the staged rungs (project-as-Rung-0 → multi-site), and an agent-automation architecture (persona × model-lineage roles, with an instrument↔subject firewall: a persona-sim is red-team / instrument-validation only, never evidence). It is a design, not a run — no results, no promotion, S17 still on hold; same-lineage draft pending independent-lineage + human review.
• Done (June 2026): full T2 grounding of the §2.7 reviewer-surfaced sources (Kahan, Bandura, Piaget, Latané, Durkheim, Rogers, Kohlberg/Rest, Buchanan & Powell) — now digest Cluster D (§9). Still deferred: external-human Round 3.

4.7 Round status

Round 4 (response / v2) complete. Develop-leg reframed from claim to testable civic design hypothesis; the decisive Round-2 failures are answered by converting them into a falsifiable test, not by re-asserting the claim. No promotion — the develop-leg is a live hold, and the §4.6 test-design memo is now drafted (design only; not run). Round 3 (external human) and Round 5 (synthesis/routing) remain reserved.

---

Standing items

• Cross-lineage provenance. Round 1 is a synthesis of Opus (baseline), GPT 5.5, and Gemini 3.1 decompositions. The three source outputs and the comparison matrix live in the steward's working thread; the divergences that shaped the synthesis are recorded in §1.6.
• The rolled-back in-lineage attempt. An earlier same-lineage attempt at this exchange (Rounds 1–2 both run by Opus) was rolled back at steward direction precisely because it auto-ran the adversarial round in the authoring lineage — the failure mode the Adversarial Review Protocol §2 default and the civic-blueprint-exchange skill §4 now guard against. Its full text is preserved out-of-tree on git branch wip/round2-rollback-snapshot-20260605; its research survives, cleaned and de-conclusioned, as the research-grounding digest.
• The biggest evidentiary gap. No source tests the comparative correlated-collapse prediction (single-authority vs. pluralist believers under the same disconfirmation) — S13's actual transferable claim and the highest-value research target if S13 is to be re-elevated.
• Dose-response gap-close (done, June 2026). The three dose-response literatures the digest originally deferred — psychological safety (Edmondson 1999), intergroup-contact theory (Allport 1954 / Pettigrew & Tropp 2006), and the trauma / productive-difficulty boundary (Tedeschi & Calhoun 2004; Frazier et al. 2009; Yerkes–Dodson 1908) — were grounded in a T2 pass and now live as digest Cluster C. A finding for S6/S7: the literatures converge on conditions over dose magnitude (safety, equal footing, shared goals, voluntariness), which is the more buildable form of S7's design gate. Still open: the matched-pluralist comparison (S13) and collective-scale dose-response (S17).
• Coverage-scan addendum (S22–S24 + §1.9 questions). Added post-synthesis by a same-lineage scan, flagged †, not tri-lineage — a completeness patch (riff prose un-lifted by the blind decompositions), not independent corroboration. See §1.6.
• Round 2 outcome (cross-lineage HOLD). Three independent lineages (GPT, Grok, Gemini), blind to each other, returned a convergent HOLD (§2): no load-bearing claim survives as a supported positive finding; S17 (cross-scale transfer) and S18 (relabel) are the decisive failures. New attacks the Round-1 self-assessment missed — Haidt post-hoc misapplication, supply-creates-demand, privilege-of-readiness, grievance-builds-architecture — are recorded in §2.5; reviewer-surfaced sources in §2.7 / digest §9. The Round 1 claim statuses are left unrewritten; the response is carried in Round 4 (§4), not by editing Round 1.
• Round 4 outcome (live hold) + next artifact. Round 4 (§4) converts the HOLD into a research program: theory conceded as relabel, novelty relocated to a testable civic design hypothesis (§4.2), and S17 recast as a test design with a pre-registered falsifier (§4.3). The §4.6 test-design memo is now drafted (June 2026) as a standalone companion: the S17 test-design memo — measurement instrument + matched comparison + the F1–F6 pre-registered falsifier set + ethics + the staged rungs (project as Rung 0) + an agent-automation architecture (persona × lineage roles; persona-sim is red-team/instrument-validation only, never evidence). It is design only — not run, no results, no promotion, S17 still on hold; like the digest it is not registered in _EXCHANGE_INDEX.md. Done (June 2026): full T2 grounding of the §2.7 reviewer-surfaced sources — now digest Cluster D (§9) (most cut against or bound the claims; named relabel antecedents for S18, a competing structural account for S17). Still deferred: the external-human Round 3.
• Test-design memo hardened by a cross-lineage run (June 2026). The test-design memo was itself run through its own §7.5 pipeline — Pipeline Run 001, four blind lineages (OpenAI / xAI / Google / Moonshot) + a non-author synthesizer. Verdict REVISE (the design survives; S17 unharmed); 5 BLOCKING fixes folded into memo v0.2 (F3↔ethics split into F3a/F3b; a pre-registered §3.6 coding rubric; §3.5 preemptive-exit closed + numeric falsifier cutoffs; consent matched to the unit; Rung-0 reclassified off the S17 ladder as pipeline-validation; citation-scope fixes). Three reviewer divergences are logged open for a v0.3 pass. The run is a Rung-0 / non-evidence instance (its own adversary flagged the Rung-0 circularity — the symmetry working).
• See-import / triad integrity. This exchange uses the mirror (one moral exercise) and the triad (S19's placement target); it does not re-grade #25's M-claims or #26's T-series.
• Provenance / register. Idea layer only (riff §1.2 privacy firewall). Working claims, named uncertainty, origin-balanced, no rhetorical flourish.