Paths: File paths (references/, ../ln-*) are relative to this skill directory.

Type: L3 Worker Category: 3XX Planning

Review Research Worker

Structured research worker for validation, audit, and review flows.

Mandatory Read

MANDATORY READ: Load references/evaluation_worker_runtime_contract.md, references/evaluation_summary_contract.md, references/evaluation_research_contract.md, references/epistemic_protocol.md MANDATORY READ: Load references/researchgraph_mcp_usage.md only when the review target explicitly cites H##, G##, benchmark run IDs, or researchgraph files.

Purpose

• gather official documentation or standards
• gather MCP Ref evidence
• gather Context7 evidence when a library or framework is involved
• gather current web best-practice evidence
• emit a compact machine-readable research summary instead of prose-only notes

Runtime

Runtime family:

• evaluation-worker-runtime

Required manifest fields:

• identifier
• phase_order
• summary_kind=review-research
• operation=research

Recommended phase_order:

1. PHASE_0_CONFIG
2. PHASE_1_RESOLVE_STACK
3. PHASE_2_OFFICIAL_DOCS
4. PHASE_3_MCP_REF
5. PHASE_4_CONTEXT7
6. PHASE_5_WEB_BEST_PRACTICES
7. PHASE_6_ANTI_HALLUCINATION
8. PHASE_7_WRITE_SUMMARY
9. PHASE_8_SELF_CHECK

Workflow

Phase 0: Config

1. Load runtime manifest.
2. Resolve review target, stack hints, and output location.
3. Fail if target context is missing.

Phase 1: Resolve Stack

1. Detect language, framework, libraries, and domain.
2. Build a bounded research topic list.
3. Keep topic count small and evidence-oriented.
4. If the target explicitly cites local H/G/run IDs, do a read-only researchgraph preflight to capture local context before external research. Do not replace official-doc, MCP Ref, Context7, or current-web lanes with local graph evidence.

Phase 2: Official Docs

1. Read official docs or standards first.
2. Record source URLs and the exact topic each source supports.

Phase 3: MCP Ref

1. Query MCP Ref for the same bounded topics.
2. Prefer primary documentation over tertiary commentary.

Phase 4: Context7

1. If a library or framework is involved, resolve the Context7 library id.
2. Query only the libraries actually used by the target.
3. If no relevant library exists, record that explicitly.

Phase 5: Web Best Practices

1. Perform current web research for best practices and recent changes.
2. Use current sources, not frozen heuristics.
3. Record only evidence that changes a conclusion or adds decision value.

Phase 6: Anti-Hallucination Verification

1. Scan target artifact for factual claims across all trigger categories per epistemic_protocol.md Section B.
2. For each claim, check against research evidence gathered in Phases 2-5:

• Has MCP Ref/Context7/Web evidence → mark VERIFIED
• No tool evidence but claim is plausible → mark FROM_TRAINING
• Contradicts tool evidence → mark FLAGGED (CRITICAL)

1. This step verifies against existing research. It does NOT run new searches.
2. Include verification status in summary metadata.

Phase 7: Write Summary

Emit summary_kind=review-research.

Payload must include:

• worker=ln-311
• status
• operation=research
• warnings

Prefer these fields when available:

• findings
• metrics.research_sources
• metrics.anti_hallucination_status (VERIFIED | FLAGGED)
• metrics.flagged_claims_count
• artifact_path
• report_path
• metadata

Phase 8: Self-Check

1. Verify all four research lanes were attempted.
2. Verify anti-hallucination verification was executed.
3. Verify skipped lanes are justified in machine-readable form.
4. Record pass=true only after the summary is written.

Definition of Done

• [ ] Official-doc evidence recorded
• [ ] MCP Ref evidence recorded
• [ ] Context7 evidence recorded or justified as not applicable
• [ ] Current web best-practice evidence recorded
• [ ] Anti-hallucination verification executed (claims marked VERIFIED/FROM_TRAINING/FLAGGED)
• [ ] review-research summary written
• [ ] Self-check passed

Version: 1.0.0 Last Updated: 2026-04-10