gongshichen/luminacode/evidence-matrix

Build an evidence matrix that can support a final report.

Inputs:

• Research contract.
• Source notes.
• sources.jsonl.
• Existing evidence-matrix.jsonl.
• Conflicts or gate findings.

Procedure:

1. Enumerate answer claims

• Extract candidate claims from source notes.
• Group claims by subquestion.
• Remove claims that do not answer the contract.

1. Link sources

• Each evidence item must reference one or more source IDs.
• Use multiple independent sources for major claims when possible.
• Use only sources whose sources.jsonl record has retrieval_status of retrieved or partially_retrieved and claim_support_allowed: true.
• Exclude sources marked inaccessible, claim_support_allowed: false, metadata-only, snippet-only, or fetch_failed.

1. Classify support

• direct empirical result.
• benchmark/dataset evidence.
• theoretical argument.
• official statement/specification.
• expert opinion.
• indirect/contextual evidence.

1. Assess confidence

• high: direct, credible, recent, and corroborated.
• medium: credible but limited or single-source.
• low: weak, old, indirect, or contested.

1. Capture conflicts

• Record conflicting evidence and why it differs.
• Do not hide inconvenient sources.

1. Identify gaps

• Missing source class.
• Missing recent evidence.
• Missing primary evidence.
• Unsupported final-report claim.

Output JSONL schema:

{"claim_id":"C001","subquestion":"...","claim":"...","source_ids":["S001","S002"],"support_type":"direct empirical result","confidence":"high|medium|low","evidence_summary":"...","limitations":"...","conflicts":["..."],"report_recommendation":"use|caveat|exclude"}

Also update conflicts-and-limitations.md with:

• Strongest conflicts.
• Evidence gaps.
• How these affect final recommendations.

Completion criteria:

• Report Writer can cite every key claim from the matrix.
• QA can verify source IDs.
• Reviewer can judge whether the method is adequate.
• No evidence item depends on a source that Source Reader did not retrieve.