Benchmark Review Candidate

Benchmark Review Candidate

Use this skill when a candidate revision needs a formal promotion review record rather than a casual thumbs-up.

It covers promotion-readiness review. It does not invent the candidate, rerun the evaluation, or silently upgrade a change because it sounds smart.

For shared benchmark framework rules — capability registry, corpus discovery, corpus adapter contract, lifecycle boundaries, and run-record finalization conventions — defer to benchmark-core.

Purpose

Use it to:

• review candidate changes against the repository's promotion gate
• compare target-suite gain against protected-suite regression risk
• check invariance and robustness stability
• verify mutation-surface compliance
• block official adoption when governance evidence is missing
• require finalized rerun evidence when direct run-record citations are part of the package
• return a canonical candidate review record

Use this skill when

Typical requests:

• reviewing a candidate_revision before promotion
• deciding whether a candidate should be hold, reject, revert, promote-to-candidate, or promote-to-official
• formalizing a promotion review after offline reruns are complete
• checking whether a new benchmark case can be adopted officially

Trigger examples

• "Help me review whether this candidate should be promoted"
• "Write a candidate gate review"
• "Check whether the protected suites regressed"
• "Assess whether this new benchmark case is ready for official adoption"

Do not use this skill when

Do not use this skill when:

• the candidate itself still needs to be proposed
• benchmark reruns have not happened and the evidence package is empty
• the evidence still exists only as half-finished scaffold placeholders that should first be finalized by benchmark-judge-run
• the task is to cluster failures or propose a wording fix
• the request is only to summarize a change informally with no gate decision

Pattern

• Primary pattern: Reviewer
• Secondary pattern: Tool Wrapper

Why this fit works:

• the core job is to evaluate evidence against explicit gate criteria
• it packages repository-specific mutation-surface and promotion rules
• a Generator-only framing would weaken the gate posture and make it too easy to drift into advocacy

Inputs

Collect or infer these inputs when you can:

• candidate ID and candidate class
• mutation-surface classification
• intended gain and observed gain
• target suite and protected suites
• judged run evidence and rerun evidence
• related failure clusters
• invariance and robustness notes
• hallucination / unsafe-compliance observations
• curator or human review state for new benchmark cases when relevant

If the package is incomplete, prefer an explicit hold posture over speculative approval.

Workflow

Use the same sequence each time.

1. Confirm candidate identity and mutation surface

Before anything else, confirm:

• candidate class
• mutation surface
• whether the surface is allowed, conditionally allowed, or disallowed

Do not review a candidate as promotable if it is quietly rewriting a disallowed surface.

2. Review the evidence package

Gather:

• target-suite results
• protected-suite results
• invariance evidence
• robustness evidence
• cluster context when relevant
• governance evidence for official adoption

Do not treat missing evidence as implied success.

If direct rerun citations still point to scaffold placeholders instead of finalized records from benchmark-judge-run, treat them as incomplete evidence and request finalization before relying on them as gate evidence.

3. Run the gate sequence

Review in this order:

• mutation-surface sanity
• scope sanity
• acceptance value
• regression protection
• invariance stability
• robustness safety
• governance fit

If one gate fails clearly, do not bury it under nicer notes from another gate.

4. Decide the outcome explicitly

Choose exactly one:

• promote-to-candidate
• promote-to-official
• hold
• reject
• revert

Local improvement is not enough if protected suites, invariance, or robustness degrade materially.

5. Return the canonical review record

Use assets/candidate-review-record-template.md for the final output skeleton.

Output contract

Return the result using assets/candidate-review-record-template.md in the same section order.

Output rules:

• intended gain and observed gain must both be explicit
• mutation-surface classification must be explicit
• target and protected suites must both be named
• acceptance, regression, invariance, robustness, and governance must all be visible
• new benchmark-case official adoption must be blocked when curator or human review is missing

Companion files

• assets/candidate-review-record-template.md — final promotion-review output structure
• references/gate-sources.md — canonical gate rules and decision guidance
• ../benchmark-core/references/benchmark-skill-lifecycle.md — shared handoff map for benchmark skills

Validation

A good result should satisfy all of the following:

• one explicit decision is made
• mutation-surface classification is present
• target-suite gain is assessed explicitly
• protected-suite regression risk is assessed explicitly
• invariance and robustness are reviewed separately
• governance status for official adoption is explicit
• follow-up actions are concrete enough for another role to continue
• direct evidence citations are stable enough that another role can follow them without reconstructing scaffold state

Common pitfalls

• approving a candidate because one benchmark improved
• ignoring protected-suite regressions
• collapsing robustness and invariance into a vague "seems okay"
• forgetting the governance requirement for new benchmark cases
• treating a blocked mutation surface as a minor technicality

Completion rule

This skill is complete when the agent has:

• reviewed the candidate against the repository gate criteria
• recorded an explicit promotion decision
• documented target and protected suite outcomes
• documented invariance, robustness, and governance findings
• returned a reusable candidate review record