---

type: adr id: 0056-adversarial-self-review-completion-discipline status: accepted created: 2026-06-08 updated: 2026-06-08 supersedes: superseded_by:

ADR-0056: Adversarial self-review is a mandatory completion discipline

Context

corpus has two relevant mechanisms, and a gap between them:

• The ## Self-review block on a task (docs/passes/implement.md, the task template) — but it is a checklist: "did I do only this step; preserve semantics; map every claim to evidence." A checklist catches omissions; it does not catch the plausible-but-wrong — the confident green summary that is simply incorrect, which is the dominant agent failure mode.
• The persona-skeptic profile (refute-by-default) — but it is scoped to independent review: its own card says "Never judge a change you authored." So the one stance built to catch the plausible-but-wrong is explicitly withheld from the actor's own work.

The result: nothing in the framework requires an implementer to adversarially review their own output before declaring it done. Self-critique measurably improves agent output [REFLEXION], and the implementer is precisely the actor most prone to over-trusting their own result — yet the discipline was absent. (Raised in dogfooding: the operator asked that finishing any work include an adversarial self-review as the skeptic.)

Decision

Adversarial self-review is a mandatory completion discipline. Before any work is marked done, the actor MUST adopt the skeptic stance over their own output — refute-by-default — and record it in ## Self-review: actively try to make each completion claim fail (re-run the bound proofs from a clean state; hunt the unexercised path — edge/error/concurrency — especially for RISK high|critical; check for scope creep and silent semantic drift; treat your own confident prose as a confession, not a proof), then fix what it surfaces.

It is necessary but NOT sufficient — and it is NOT the gate. This reconciles with implementer ≠ reviewer (the self-preference hazard, docs/passes/review.md): adversarial self-review produces fixes + a recorded self-critique, never a gate verdict. A self-issued PASS remains inadmissible; the independent review step still renders the merge gate. The skeptic persona's "never judge a change you authored" continues to bar the implementer from issuing the verdict — it does not excuse them from attacking their own work first. Self-review is the cheap pre-handoff pass that catches your own holes before an independent reviewer (or the gate) has to; the two are complementary, not substitutes.

Scope: every completed deliverable. For implement/fix, refute the proofs and the scope/semantics boundary. For an authoring/analysis deliverable (spec, audit, research, bug-report, or a framework change like this one), the adversarial pass is try to refute your own obligations / findings / claims before declaring the artifact done.

Alternatives considered

Alternative	Why rejected
Keep the checklist ## Self-review	A checklist catches omissions, not the plausible-but-wrong; that is exactly what the skeptic stance — and [REFLEXION]-style self-critique — exists to catch.
Let the skeptic persona self-judge the gate on its own work	Breaks implementer ≠ reviewer and the self-preference hazard. A self-issued verdict is not independent. Self-review must produce fixes + a critique, never a verdict.
A brand-new step in the 9-step flow	Out of scope and unnecessary — it strengthens an existing block (## Self-review) and an existing profile (persona-skeptic); it changes no closed set.

Consequences

• Positive: the plausible-but-wrong is attacked at the cheapest point — by the author, before handoff; the discipline is standing, not ad hoc; the skeptic stance is no longer withheld from one's own work.
• Negative: one more pass on every task — proportionate, since a defect caught at self-review is the cheapest defect to fix.
• Neutral: no closed set, grammar, step, verdict, or artifact change; implementer ≠ reviewer and the independent gate are untouched (explicitly reaffirmed).

Status

Accepted (v0.1). The ## Self-review strengthening, the task-template prompts, the persona-skeptic + implement-and-verify updates, and the universal-rule addition are this change.

Affected obligations / constraints

• Refines: the implement step (docs/passes/implement.md ## Self-review), persona-skeptic, the implement-and-verify skill, and the universal working rules (AGENTS.md).
• Reaffirms (does not change): implementer ≠ reviewer and the independent merge gate (docs/passes/review.md).
• Grounded by: [REFLEXION] (self-critique improves output) + the self-preference hazard already cited in review.
• Does NOT change: any closed set, the SOL grammar, the nine steps, the verdict model, or the artifact set.

Ledger note (2026-06-11): refined by ADR-0064 (self-review carried inside the implement-task guide).