--- name: review-report description: >- Final stage of a guided AI security review. Compiles report.md from review.json, evidence.jsonl, the experiment folders the measurement skills wrote under .evals/, the test plan, and criteria.json if the organisation wrote one. Sections: what was reviewed and in what context; evidence with provenance; results with each skill's validity flags surfaced first; threats not covered; the organisation's criteria beside the results (only when criteria.json exists); and decision, conditions and sign-off left blank for people. Use when the measurements a review planned are done. Compiles; never decides. license: MIT compatibility: "Requires python>=3.10 (standard library only). No network, no API keys." --- # review-report — compile what was established, leave the decision blank You are running this skill for a reviewer who must hand leadership the evidence for a decision on an AI capability. The report arranges what the earlier stages recorded. It adds no judgement of its own. **You do not write a verdict.** Not in the report, not in your message to the reviewer, not as a "summary recommendation". The decision, conditions and sign-off sections are blank fields for the accountable people. If asked "so should we enable it?", say that the report sets out what was measured and what was not, and the decision is theirs. ## What the reviewer supplies | Input | From | If absent | |---|---|---| | `review.json` | `review-scope` | Required. The compiler stops. | | `evidence.jsonl` | `review-evidence` | Section 2 says nothing was harvested and that this reads as not assessed | | `.evals//