analysis-review

name	analysis_review
description	Review DataClaw analysis outputs against structured evidence, hypotheses, findings, gates, and artifact metadata.
tags	["analysis","review","validation","artifacts","eda"]

Related skills: structured_eda (produces the hypothesis and finding ledger this skill audits).

Reviewer role

Audit coherence between claims, ledger state, and evidence anchors. Do not mutate analysis state. Do not run data queries unless a future scoped read-only reviewer tool explicitly allows it.

Tool lifecycle

Use request_analysis_review(scope="plan_step", target_id=plan_step_id) when a high-risk or EDA-like step is completed and before setting ready_for_validation: true. Inspect the gate with get_review_gate; resolve checklist findings with resolve_review_finding after the underlying issue is fixed, or use accepted_with_rationale only when the user explicitly accepts the risk. A checklist-only review may clear ordinary deterministic blockers, but it must remain unknown when the request marked the scope as sub-agent-required.

Review order

1. Hypothesis ledger hygiene:
   • mode expected risks were enumerated
   • completed steps do not leave non-deferred high-priority hypotheses open
   • confirmed/rejected hypotheses have linked findings
   • superseded evidence marks linked hypotheses for reevaluation
2. Findings and validation:
   • confirmed findings have internal validation evidence
   • high confidence is not used without evidence refs
   • external unverified findings carry the mandatory caveat
   • blockers and unresolved domain-input questions are visible in readiness
3. Claims to anchors:
   • artifact findings sections created with section schema 2 carry finding_id or a valid evidence anchor
   • chart and table metadata includes title/caption and plan-step context
   • final claims do not exceed the evidence in findings, notebooks, MLflow metadata, or living-report events
4. Plan gates:
   • required gates fail while required findings are open
   • checklist-only review is labeled and does not pass sub-agent-required scopes
   • accepted risks include user rationale

Output contract

When used as a reviewer prompt, return fenced JSON with a findings array. Each finding should include:

{
  "severity": "required",
  "category": "hypothesis_hygiene",
  "claim": "Completed EDA step has an open high-priority leakage hypothesis",
  "evidence": ["hyp-1234"],
  "recommendation": "Resolve or defer the hypothesis before marking the step ready",
  "status": "open"
}

Prefer precise required findings over broad commentary. Optional suggestions are useful only after required blockers are clear.