ADR-0122 — Revolver Review: a bounded, named form of the lead-orchestrated review panel

Context

ADR-0099 established the review model: a review lead chooses at least three distinct lenses, runs them as independent subreviewers that return findings + evidence only, reconciles into one packet, and the human owns the verdict. The suspec-works roadmap (#77/#78/#79/#85) proposes Revolver Review — a named strategy that adds a "stance wheel", a candidate-finding schema, a deterministic-check/context-slicing policy, and a fixed configuration (9 stances / 3 agents / 3-per-agent / 2 rounds).

A July 2026 deep-research pass verifying that roadmap against primary sources found the direction sound but four things needing correction before it can be leaned on: the diversity payoff is modest and about ordering — a few genuinely distinct lenses beat more identical copies — not the claimed large ratio, and its matched-count delta is small ([DIVSCALE]); transfer to code review is unconfirmed (the supporting evidence is small models on QA/math); the invented counts are unproven; and one framing — Suspec runs the deterministic checks as a pre-gate — collides with the reconcile-only boundary (ADR-0086 Decision 5, ADR-0077 D8). This ADR names Revolver as the bounded form of ADR-0099's panel and records the review-packet capture it needs — additively, without changing the frozen format.

Decision

Revolver Review is the bounded, named form of the ADR-0099 lead-orchestrated independent-lens panel — it adds no new review authority.

1. The flow is ADR-0099's, bounded. A lead picks a few distinct lenses from a menu by the change's risk; same-round lenses run blind and isolated and draft before they compare (reading a peer's raw output induces conformity — [CONSENSUSCOST]); the lead reconciles into one report; the human owns the verdict (ADR-0077 D8). A lens reviewer never renders the status or the suggested decision. The cross-round exchange format — total isolation versus an adjudicated summary — is open: the tested remedy is isolation ([CONSENSUSCOST]), so same-round isolation is the load-bearing rule and any cross-round summary is a conservative design choice, not an evidence-backed one. Level: convention/checklist — nothing enforces it.

2. The stance wheel is a lens MENU; counts are measured defaults, not law. No stance/agent/round count is canon — not 9/3/3/2, not any number. The strategy stops adding lenses/rounds when the marginal unique accepted findings dry up (plus hard budget/round caps), never on silence or agreement. Any count or cheap-tier default is validated on the review-gate benchmark by marginal unique-accepted-finding yield and cost-versus-unique-yield; the diversity payoff is modest and about ordering ([DIVSCALE]), and code-review transfer is flagged unconfirmed. Level: convention.

3. Candidate-finding adjudication is captured additively. Lens output enters as a candidate and the lead moves it to accepted / rejected / duplicate / unverified / blocked. This is an optional, additive capture in the review packet — an opt-in section that leaves the frozen requirement-coverage table and the Pass/Fail/Unverified/Blocked verdict enum unchanged (the ADR-0060 Decision 4 / ADR-0099 additive-amendment pattern; the review packet format does not change). An accepted finding carries concrete evidence; reviewer confidence is non-evidence; deduplication feeds the unique-finding count the stop rule reads; a review is never auto-closed on agent agreement (agreement is not a correctness signal — [CORRELATED]). Level: convention (the kit review template carries the optional capture).

4. Deterministic checks are reconciled, not run. The reviewer or worker runs the project's build/test/lint commands; Revolver reconciles the pasted, already-run output and never spawns the project's commands (ADR-0086 Decision 5, ADR-0077 D8). The lever is reliable, actionable structured evidence over free-form critique (ADR-0121; [CRITIC-TOOL], [FLIPFLOP]) — a surfaced, interpretable result, not a raw trace. Level: convention.

5. Per-lens cost tiers are opt-in and owned by the model-routing convention (a separate decision): blind lenses default cheaper, the reconciling lead and high-risk lenses (security, architecture) stronger, resolved by the runner — no shipped model defaults, and model size is never a quality proxy. A mixed-tier panel also decorrelates the reviewers. Level: convention.

Alternatives considered

Consequences

• No verdict-model change and no format re-freeze. Pass/Fail/Unverified/Blocked and the coverage table stand; the candidate-finding section is additive and opt-in.
• adversarial-review is unchanged — each Revolver lens runs it; no rename, no absorption.
• The counts stay honest. The public strategy carries no count-bearing prose asserted as proven (ADR-0117); the numbers live in the measurement, not the doctrine.
• The review-gate benchmark gains a Revolver measurement track (marginal unique-yield per lens/round; cost-vs-yield per lens/tier) — a reopen of its v1 list, not new standing infrastructure.
• The universal revolver-review skill (suspec-skills) carries the strategy framework-free and citation-free (ADR-0113); the evidence lives here.

Status

Accepted (2026-07-02). Refines / extends ADR-0099 (names and bounds the panel). Honors ADR-0077 D8 and ADR-0086 Decision 5 (reconcile-only), ADR-0121 (evidence-gating), ADR-0060 Decision 4 (additive packet amendment), and ADR-0117 (no count-bearing prose). Relates ADR-0095, ADR-0056, and ADR-0119. The per-lens cost-tier convention is a separate decision (the model-routing wave).

Forward note (2026-07-03) — union-and-reconcile, never majority-filter. The lead's reconcile unions every lens finding and adjudicates each on evidence; agreement or vote-count never selects which findings survive, and a vote applies only to a contested rejection, refutation-first. Clarified after measured evidence that consensus-style aggregation dilutes a team below its best member (up to 41.1%, integrative compromise — [EXPERTSBACK]); recorded in SPEC-revolver-union (family workspace). The body above is unchanged.

Superseded (2026-07-04) by ADR-0132. The bounded parallel-panel form recorded here is replaced by the rotating adversarial refine-loop — a 6–9 stance pool fired 3 at a time, the orchestrator fixing between rounds and the next trio re-reviewing the revised state, rotating one stance per round, up to 3 cycles. The union-not-vote reconcile and the cheap-varied-model routing carry forward; the panel-over-a-frozen-diff mechanism does not. The body above is unchanged, per the immutable-ledger rule.