Evaluator Artifact Contract
| Language: πΊπΈ English | π°π· νκ΅μ΄ |
This artifact contract applies to verification-complexity-level-3-plus work, intake of external agent-governance logic, and RAG/evaluator-driven improvement candidates. Its purpose is not to produce more reports; it is to preserve promotable evidence together with rejected candidates.
Contents
Applicability
- verification-complexity-level-3-plus work
- Intake of external agent-governance logic
- RAG or evaluator output that proposes playbook, rule, skill, or hook changes
- Promotion decisions before installed assets change
Artifact Set
scenario.json: records the verification scenario, input conditions, non-goals, and forbidden surface.trace.json: records evaluator execution trace, read-only status, accessed sources, and command summaries.report.json: records metrics, findings, pass/fail state, and uncertainty.candidate-playbook.md: explains the candidate procedure for a human reviewer.verifier-result.json: records verifier judgment, accepted/rejected state, and rejected candidate list.
Promotion Rules
- A read-only evaluator pass must not mutate installed assets.
- Do not promote
candidate-playbook.mdwithout an accepted verifier-result (verifier-result.json). - At least one rejected candidate must exist so the verifier has proven it can say no.
- If
verifier-result.jsonhasaccepted: false, stop installed asset changes. - This artifact contract is for high-risk work. Do not use it to manufacture reports for low-complexity tasks.
Minimum verifier-result.json Shape
{
"accepted": false,
"reason": "insufficient evidence",
"accepted_candidates": [],
"rejected_candidates": [
{
"id": "candidate-001",
"reason": "mutates installed assets during read-only pass"
}
]
}
Forbidden
- Mutating installed assets during a read-only pass
- Formal verifier flow that accepts every candidate without rejected candidates
- Promoting evaluator output directly into a skill, hook, or rule without evidence
- Creating report loops whose purpose is only to create reports