ADR-0132 — Revolver Review is a rotating adversarial refine-loop (supersedes ADR-0122; refines ADR-0124; removes the adversarial-review skill)

Context

ADR-0122 named Revolver Review a bounded parallel panel: pick a few lenses, run them blind at once over one frozen diff, reconcile once, hand a human the findings. That is a review gate. It does not match the mechanism the name implies — a revolver's cylinder rotates, chambers fire a few at a time, and each shot lands on the current target. Two gaps followed from the panel framing: it reviews the original diff and is therefore structurally blind to the defects a fix introduces, and it never converges the work — it only reports.

A July 2026 evidence pass (inline; a few web searches, not a fan-out) grounded the corrected mechanism. The moves are well-supported; the specific numbers are evidence-consistent defaults, not proven optima — recorded honestly so no reader mistakes a tuned default for a measured constant.

• Re-review the revised state. Iterative fixing routinely introduces new defects; only re-reviewing the post-fix code catches them — a one-shot panel cannot. (Iterative-repair evidence; the regression hazard is the load-bearing point.)
• External feedback, not self-correction. Intrinsic self-correction fails and can degrade output ([[NOSELFCORRECT]]); fresh, independent reviewers on the revised state are the external signal ([[CCR]]) — so the party that applied a fix is never a reviewer of it.
• Blind within a round, adjudicated between rounds. Reading peers' raw drafts induces sycophancy ([[CONSENSUSCOST]], [[FLIPFLOP]]); a later round receives the orchestrator's reconciled/fixed state, never another reviewer's raw draft.
• Distinct perspectives, cheap and varied models. Coverage comes from perspective diversity, not reviewer count ([[PBR]]); correlated reviewers ≈ one reviewer, and diverse signals beat more compute ([[ENSEMBLEDIV]], [[DIVSCALE]]) — so cheaper varied models are the default, not a compromise.
• Union, never vote. Consensus aggregation drags a panel below its best member ([[EXPERTSBACK]]).
• Cap the loop. Iterative gains plateau fast and extra rounds add noise/over-edit risk; more review volume dilutes ([[REVBOTPR]]); a single agent can match a panel under matched compute ([[SINGLEBEATSMAS]]).

Decision

Revolver Review is a rotating, self-converging adversarial refine-loop. Level: convention.

1. The mechanism (superseding ADR-0122's bounded-panel form): a pool of 6–9 distinct stances; fire 3 reviewer subagents per round, blind and isolated; the orchestrator applies the legitimate fixes between rounds; the next round rotates one stance in and one out and reviews the revised state; continue to a full rotation (every stance fired once), then repeat for up to 3 cycles, stopping when a cycle surfaces no new accepted finding.
2. The numbers are evidence-consistent defaults, to measure and tune — not constants. The moves (revised-state, fresh-blind reviewers, fix-between, union, cheap-varied, capped cycles) are well-grounded above. The magnitudes — pool of 6–9, trio of 3, one-in-one-out, ≤3 cycles — sit in the direction the evidence points (small diverse batches capture most of the gain; ~10 review dimensions give coverage; iteration plateaus fast) but are not pinned to a verified optimum. Treat them as defaults a team measures against marginal unique-accepted yield, per ADR-0124. (Honesty marker: a widely-summarized "k=3 optimum" figure did not survive primary verification and is deliberately not asserted — see [[ENSEMBLEDIV]].)
3. Cost-tiering defaults to cheap; the strong tier is the opt-in (refines ADR-0124's opt-in framing): reviewer chambers run on lower-end, varied models by default so the loop is affordable to run across cycles; escalating a specific stance (security, architecture) or the reconciling orchestrator to a stronger model is the deliberate opt-in — never the whole panel by default, never a silent inherit of an expensive session model. Level: convention.
4. The adversarial-review skill is removed; its discipline is consumed into revolver-review. A solo refute-by-default review with no rotating loop adds nothing a harness lacks out of the box; its value — refute-by-default, re-run/reconcile the evidence, cite file:line, keep effective false positives low, the adversarial questions — is inlined into revolver-review and injected into each reviewer subagent's prompt alongside its stance, so the skill is self-contained and installs alone (no skill depends on another). The adversarial self-review discipline (ADR-0056) is unaffected — it remains canon and is carried in the kit AGENTS.md floor and the suspec-reviewer agent, which already self-contain it.

Consequences

• revolver-review (suspec-skills) is rewritten to this mechanism, self-contained. adversarial-review is deleted; the catalog's dozen references repoint to revolver-review; suspec-reviewer drops its optional see-also. No contract change, no check; convention-level (ADR-0063), no count-bearing prose in product tiers (ADR-0117 — the magnitudes live here as defaults, with their honesty caveat).
• Independence and the evidence gate are untouched: reviewers read-only, the orchestrator fixes, a human owns the ship call (ADR-0119, ADR-0121).
• The DP-8 panel measurement (suspec-works) was taken on the bounded-panel form; a fresh measurement of the rotating loop (marginal unique-accepted yield per cycle, cost per cycle) is the honest follow-up — the numbers in Decision 2 are defaults until it lands.

Status

Accepted (2026-07-04). Supersedes ADR-0122 (the bounded-panel form → the rotating refine-loop). Refines ADR-0124 (cost-tiering defaults to cheap; strong is the opt-in). Removes the adversarial-review skill, consuming its discipline. Honors ADR-0056/ADR-0119/ADR-0121/ADR-0063/ADR-0117. Grounded by [[SELFCORRECT]], [[NOSELFCORRECT]], [[CCR]], [[PBR]], [[ENSEMBLEDIV]], [[DIVSCALE]], [[CONSENSUS]], [[EXPERTSBACK]], [[BIGGERNOTBETTER]], [[SINGLEBEATSMAS]], [[REVBOTPR]], [[CONSENSUSCOST]], [[FLIPFLOP]], [[SELFREVIEW-MOD]].

Refined (2026-07-04) by ADR-0133. The mechanism default is updated from fire 3 at a time over a 6–9 pool with one-in-one-out rotation to one reviewer at a time over a pool of at least 6 distinct stances — even coverage by construction, no fixed upper limit. ADR-0133 records the examine-don't-ruminate basis and labels the single-vs-panel choice a reasoned default with an A/B measurement queued. The adversarial-review skill removal and the cheap-default tiering recorded above are unchanged.