Use after Breaker ACCEPT and before Reliability scoring — gates completion on evidence that smoke steps actually pass.
Run this phase after Breaker has accepted and before Reliability scoring. No change is scored until there is evidence it works.
Every claim of success requires fresh output from this session proving it. If a smoke step was not run, it did not pass.
HEALTH_CHECK from the Builder's output.SKIPPED — <reason>.## VERIFICATION
### Passed
- [ ] `<command>` — <outcome summary>
### Skipped (with justification)
- [ ] `<step>` — SKIPPED: <reason it cannot be run here>
### Failed
- [ ] `<command>` — <failure output>
If any smoke step fails, do not proceed to Reliability scoring. Return to Builder with the failure output as a new rejection item. This counts toward the 3-cycle rejection limit.
If every smoke step is skipped, the Reliability score must reflect this: a change with zero verified steps cannot score above 6/10 regardless of code quality.
| Shortcut | Why it's wrong |
|---|---|
| "Tests should pass" without running them | Speculation is not evidence |
| Running a subset and claiming full pass | Untested paths are unknown paths |
| Marking steps as "verified" from prior context | Prior context may be stale; re-run in this session |
| Skipping verification because "the change is simple" | Simple changes break production too |