# One-shot evidence completeness > Applies when: a failed attempt is not safely retryable (consumptive action). An action is **consumptive** when a failed attempt is not safely equivalent to a normal rerun. This includes commands limited to one attempt, commands whose retry needs fresh human authorization, actions that spend external quota or mutate privileged/external state so a rerun is not equivalent, and actions whose cleanup can destroy the failure state needed for diagnosis. Successful cleanup does not make a consumptive action repeatable. ## Pre-attempt gate Before Vet can emit READY, and again immediately before Prove executes the action, the approved `test-plan.md` must bind all of the following: 1. **Durable retention:** an operator-controlled evidence artifact outside the disposable runtime/cleanup tree, created before the first side effect, written durably before cleanup, least-privilege, and bounded by schema, size, and cardinality. 2. **Trust-safe diagnostics:** known semantic values use the normal validator; unknown but lexically well-formed non-secret values survive in bounded sanitized fields; malformed, hostile, or secret-bearing values become fixed reason codes rather than retained raw input. 3. **Terminal completeness:** every success, nonzero exit, rejection, timeout, signal, and cleanup failure either names the retained artifact or proves that no diagnostic state exists. Failure retention preserves the original safe failure family and cause through clean convergence. 4. **Discriminating proof:** fixtures cover success, a known failure, an unknown well-formed failure, malformed/hostile input, and cleanup after failure. They prove cleanup cannot delete or overwrite the retained failure evidence. 5. **Causal actionability:** every failure record includes a stable non-secret `boundary_id` whose finite map is injective: one retained fingerprint identifies one actionable failure seam and correction class. Broad operation/cause labels are insufficient when multiple emit sites can produce them. `test-plan.md` enumerates every emit site, its boundary ID, expected retained relation, and offline decision. 6. **Collision proof:** inject a failure at every mapped seam and require its exact boundary ID. Execute a negative mutant that aliases two seams to one retained fingerprint and prove the validator/reviewer rejects it. 7. **Recovery sufficiency:** the retained bounded evidence is enough to choose an offline correction or a truthful terminal classification without consuming another attempt. Missing or stale evidence is an agent-owned technical plan gap: Vet returns `NEEDS REPLAN`, and Prove returns to Vet inline without executing the action. Never weaken the trust validator or spend the attempt merely to discover what the retention design should have preserved. ## Failure handling After a consumptive action fails, its retained artifact is the reproduction input. Do not rerun the action during triage. Keep two budgets separate: - **Action authorization:** the failed execution consumes only the authorization for that consumptive execution. Zero remaining action attempts prohibits another real execution; it does not exhaust offline diagnosis or correction. - **Causal-fingerprint recovery:** when the retained artifact supplies a new Critical/Important failed invariant, the controlling caller immediately runs bounded offline triage, repair, fixtures, and narrow Vet in the same invocation. Count only no-progress corrections of that exact fingerprint under `afk-hitl.md`; do not stop merely because the action authorization was consumed. Cold resume does not make that fingerprint old or exhausted. Derive its offline no-progress count from `drift.md` and `evidence.md`; while the count is below the cap, resume recovery even if a prior writer stored `blocked` / `Next step: none`. That terminal cursor is valid only for a human/safety gate, an actually exhausted fingerprint, or a proven absence of any safe retention/amplification design. A new real attempt is normally admissible only after the affected plan and fixtures are re-vetted, the failure condition is shown changed, and any required fresh authorization is obtained. Stop at that authorization boundary; never infer it from successful offline repair. ## Diagnostic amplification If a retained artifact is absent or maps one fingerprint to multiple causal boundaries, do not guess a runtime correction. The fact that evidence from the past attempt is irretrievable does not prove that a safe future evidence-acquisition design is unavailable. When an in-scope trusted seam can add the missing stable discriminator, classify the ambiguity as an agent-owned **diagnostic-amplification plan gap**. Without executing the action, repair the diagnostic schema, finite boundary map, per-seam fault fixtures, cleanup-survival proof, and collision mutant; then run narrow Vet. Once READY, stop for fresh authorization before exactly one diagnostic-amplification attempt bound to that vetted design and artifact identity. This exception does not claim the runtime failure condition changed: the independently proven change is the evidence-acquisition invariant, and the attempt's acceptance signal is the promised unique retained boundary (or action success). Terminal `Next: none` is valid only when no safe in-scope diagnostic-amplification seam exists, the required change is human/risk/scope owned, or the exact evidence-gap fingerprint exhausts bounded recovery. A missing old artifact alone is not terminal, and an amplification attempt never inherits authorization from the failed action.