Lists multiple specific concrete actions: 'recording baselines', 'applying hypothesis patches', 'comparing metrics', 'protecting regression guards', and 'deciding keep, discard, rollback, or block'. These are detailed, actionable steps.

Clearly answers both 'what' (run bounded automated experiment iterations with specific steps) and 'when' (explicit 'Use when' clause covering '$autoresearch' naming, evidence-backed research needs, metric tracking, or safe optimisation loops).

Includes some useful trigger terms like '$autoresearch', 'metric tracking', 'evidence-backed research', and 'optimisation loops', but many are somewhat technical. Natural user phrases like 'run experiment', 'A/B test', 'try and measure', or 'benchmark' are missing. The '$autoresearch' keyword is a specific command trigger but not something users would naturally say.

The combination of experiment iteration loops, hypothesis patches, regression guards, and the specific '$autoresearch' trigger creates a very distinct niche. This is unlikely to conflict with general coding, testing, or research skills.

Generally efficient and avoids explaining basic concepts, but some sections are somewhat verbose or redundant (e.g., 'Avoid', 'Anti-Patterns', and 'Gotchas' overlap in intent; the Discovery Interview section adds little actionable value). Several bullet points could be tightened.

Provides concrete examples (ledger YAML, shell iteration example, decision criteria), but much of the guidance remains at the policy/principle level rather than executable steps. The workflow steps mix concrete commands with abstract directives like 'Define parser contract, guard command, held-out checks' without showing how.

The 8-step workflow is clearly sequenced with explicit validation checkpoints (baseline first, verify + guard at each iteration, fail-fast at first failed gate, repair before proceeding). Feedback loops are well-defined: keep/discard/crash/block decisions with evidence, and the repair behavior section adds error recovery guidance.

The Progressive Disclosure section at the end references several external files (references/autoresearch-project.md, references/contract.yaml, references/evals.yaml, references/task-profile.json, references/discovery-interview.md), which is good structure. However, no bundle files are provided to verify these exist, and the main body still contains substantial inline detail that could be offloaded. The references are listed but not clearly signaled with context about when to consult each one.