| name | verify-evidence |
| description | Verification checklist for AC, test evidence, stuck-loop detection, AFK review. Use when lead needs independent evidence for non-trivial changes or ambiguous results. |
Load this skill when validation-lead or engineering-lead needs independent evidence. Do not load for trivial docs/config changes.
When to Use
- Non-trivial changes (behavior, security, data)
- AFK/unattended tasks
- Flaky or failing tests need confirmation
- Explicit acceptance criteria must be verified
- Deployment, auth, DB, or payment changes
- Reviewer needs independent evidence
When to Skip
- Trivial docs/config changes — inspect diff directly
- No acceptance criteria AND no high-risk/AFK/validation-lead evidence need exists
- Review is purely style/naming suggestions
Self-Improvement Pipeline Integration
Part of self-improvement pipeline (see ADR-0003: Skills Recategorization & Memory Protocol).
Before verifying, scan .scratch/evals/ for findings matching current change's target or category. Flag if CRITICAL/HIGH findings exist for same skill — verify fixes applied. Escalate to BLOCKED if same finding repeats 2+ times.
Memory protocol: After gate, store ov add-memory "[workflow:verify] <change>: <status>, <N> eval patterns checked".
Evidence Mapping
Map each acceptance criterion to a verifiable signal. Run commands, inspect output.
| Criterion | Command/Evidence | Status |
|-----------|------------------|--------|
| "form submits" | curl POST → 200, no console error | PASS |
| "no layout shift" | screenshot A vs B diff < 1% | PASS |
Statuses
| Status | Meaning |
|---|
VERIFIED | All criteria have PASS evidence |
FAILED | Any criterion has FAIL evidence |
INCONCLUSIVE | Verification could not fully run (missing tools, broken infra) or criterion lacks objective signal |
BLOCKED | Human checkpoint needed or stuck loop detected |
Stop Conditions
Immediately report BLOCKED:
- Destructive ops without checkpoint:
DROP TABLE, DELETE FROM, rm -rf, billing commands
- Remote irreversible state: deployment to production, DNS changes, API key rotation
- Cost >$1: operation exceeds $1 in API calls without prior approval
- Secrets/security boundary: verification step would write outside allowed paths or make unauthorized network calls
- Human decision needed: acceptance criterion cannot be mapped to verifiable signal and human judgment required
- Stuck loop: same failure pattern 3+ times with no progress
Stuck-Loop Heuristic
Check for:
- Same error message appearing 3+ times in consecutive runs
- Output size not growing across retries
- No new files or file changes after M minutes
- Builder repeatedly touching same files with no resolution
Output Format
Status: VERIFIED | FAILED | INCONCLUSIVE | BLOCKED
Evidence:
- [criterion/risk] — [command/check] — [result]
Gaps:
- [missing evidence or none]
Blocked:
- [if BLOCKED: blocker + needed input]
Rules
- No fixing in verify mode. Report evidence only. Do not implement fixes, edit files, or commit.
- Report what tools say. Do not re-interpret results optimistically. Pass = clean exit + expected output.
- If verification command is unavailable, say so. Do not simulate results.
- Keep compact. Prefer evidence paths over inline dumps. Reference log files, screenshots, test outputs.
- No secrets in output. Truncate logs >200 lines; link to evidence files.
