| name | implementation-verification |
| description | Verify branch changes against origin/main with claim-based evidence, not only test status. |
| category | validation |
| kind | atomic |
| phase | verification |
| requires_write | false |
| requires_slurm | false |
| requires_benchmark_artifacts | false |
| delegates_to | [] |
| output_schema | skill_run_summary.v1 |
Implementation Verification
Purpose
Prove that branch changes implement the stated claims versus origin/main, with evidence per behavior
rather than only a global test pass signal.
Workflow
- Set baseline:
git status --short --branch
git diff --stat origin/main...HEAD
git diff --name-only origin/main...HEAD
- Extract claims from issue/PR body, docs, and changed code/docs.
- Map each claim to a concrete evidence surface:
- scripts, configs, tests, CLI commands, or artifacts.
- Validate claim by claim:
- run minimal targeted checks first,
- then run
BASE_REF=origin/main scripts/dev/pr_ready_check.sh when broader validation is required.
- Include benchmark safety checks when relevant:
- explicitly classify fallback/degraded execution as a limitation, never as success.
- Record residual gaps where no direct proof path exists.
Guardrails
- Do not treat a passing test suite as sufficient if it does not exercise the claim.
- Do not edit unrelated code while proving current claims.
- If
origin/main is unavailable, switch to an explicit documented base and note why.
- Treat delegated worker artifacts as leads until locally checked.
status=0 only proves the wrapper
exited, not that the model returned useful findings.
- For read-only delegates, require a structured final answer in
RESULT.md; otherwise keep the
worker output at low confidence and inspect local evidence directly.
- When worker metadata says
changed_files: [], still inspect status.txt or git status --porcelain
for untracked files before using the run as implementation evidence.
Output
- Evidence matrix (
claim, proof, command/artifact, result, residual risk).
- Items proven vs unresolved.
- Recommended follow-up tickets for unproven claims.
When to use
Use this skill for the scope named in its frontmatter description and registry metadata.