| name | loop-on-verification |
| description | Use when a task has explicit, mechanically checkable acceptance criteria and the agent should iterate implement, verify, and revise until GREEN or a maximum of three iterations is reached. |
Loop On Verification
Run a bounded execution loop against acceptance criteria.
Checklist
- Read the issue, goal file, or task prompt and copy the acceptance criteria.
- Set
max_iterations = 3.
- Implement only the smallest change needed for the current criterion.
- Evaluate with a reviewer role such as
replication-verifier,
identification-reviewer, or data-reviewer.
- Record each criterion as PASS, FAIL, or UNABLE TO CHECK.
- If all pass, stop with GREEN and request human diff review.
- If a minor gap remains, revise only that gap and continue.
- If a blocker cannot be resolved within scope, stop with RED and draft a new
issue for the blocker.
Card-Krueger Example
Use this loop for a robustness task only if the goal states:
- allowed files;
- required command output;
- required tests;
- raw or synthetic data files that must remain unchanged;
- explicit caveat that synthetic data do not support a causal claim.
Output
# Verification Loop
| Iteration | Verdict | Evidence | Next Action |
| --- | --- | --- | --- |
## Final Status
## Remaining Human Checks
Constraints
- Do not loop indefinitely.
- Do not auto-merge after GREEN.
- Do not broaden scope; new findings become new issues.
