evaluation-preflight
Validate an Arena evaluation protocol before any costly or result-producing run.
SKILL.md.claude/skills/evaluation-preflight/SKILL.md
GitHubPerform a read-only preflight for Arena evaluation $0. Do not start the evaluation.
Read AGENTS.md and the governing Issue/protocol, then verify only requirements that are actually specified, including as applicable:
- execution target and required merged revision
- clean-worktree requirement
- candidate and baseline identity
- seed set, game count, duplicate layout, and worker constraints
- artifact/output location and overwrite policy
- pre-execution lock or receipt requirements
- one-shot, rerun, or result-publication restrictions
- dependency/pin identity
- estimated runtime/cost when the protocol makes it decision-relevant
Treat application/protocol-level validation as authoritative. Claude Hooks are only operator-error guards and must not be treated as the research-integrity boundary.
Report PASS, BLOCKED, or NOT APPLICABLE for each relevant item and list any command the user should run next. Do not execute the result-producing command.