gongshichen/luminacode/citation-qa-report

Audit the research artifacts for traceability.

Before reviewing:

• Call GetTeamContext.
• Use runtime contract as authoritative.
• Read required artifacts from the Lumina team runtime directory.

Procedure:

1. Load context and artifacts.
2. Validate required artifacts.
3. Parse source registry and evidence matrix.
4. Cross-check final-report source IDs.
5. Classify findings as blocking or nonblocking.
6. Write qa-report.md.
7. Call SubmitGateVerdict for citation_qa.

Checks:

1. Artifact completeness

• Required artifacts exist.
• Files are non-empty and relevant.
• Runtime artifacts are not accidentally written into project root unless requested.

1. Source registry

• sources.jsonl parses as JSONL.
• Each source has source_id, title, url or arxiv_id, source_type, retrieval_tool, retrieval_status, claim_support_allowed, and notes_path.
• Source IDs are unique.
• Sources used by evidence items are retrieved or partially_retrieved and have claim_support_allowed: true.
• retrieval_status values are only retrieved, partially_retrieved, or inaccessible.

1. Evidence matrix

• evidence-matrix.jsonl parses as JSONL.
• Each evidence item has claim, source_ids, confidence, support_type.
• Referenced source IDs exist in sources.jsonl.

1. Final report citations

• Key claims cite source IDs.
• Cited source IDs exist.
• No major claim relies only on snippets.
• final-report.md contains a readable source index and evidence index. The evidence index must map evidence IDs to source IDs so the user can audit the report from the working-directory deliverable package.

1. Retrieval quality

• Web claims have WebFetch-backed sources or documented shell/curl fallback sources after a recorded WebFetch/WebSearch failure.
• Academic paper claims have arXiv MCP or fetched paper notes when applicable.
• No evidence item is marked as supported by model memory, training knowledge, snippets, or unretrieved metadata.
• Inaccessible sources are clearly excluded from final-report claim support.

Blocking findings:

• Any final-report key claim without a valid source ID.
• Any cited source ID missing from sources.jsonl.
• Any claim supported only by search snippets, model training knowledge, memory, unretrieved metadata, or undocumented fallback retrieval.
• Any inaccessible source used as if it were retrieved evidence.
• Any evidence item citing a source whose retrieval_status is missing, inaccessible, fetch_failed, or whose claim_support_allowed is false.
• Any source record using invalid retrieval status aliases such as success, fail, fetched, or fetch_failed.
• Any Search Strategist candidate list being treated as sources.jsonl without Source Reader retrieval notes.
• Missing sources.jsonl, evidence-matrix.jsonl, or final-report.md when required by the contract.
• Missing source/evidence index in final-report.md.

Output:

Write qa-report.md:

# Citation QA Report
 
## Verdict
pass|fail|not_applicable
 
## Artifact Checks
| Artifact | Status | Evidence |
 
## Source Registry Checks
...
 
## Evidence Matrix Checks
...
 
## Citation Checks
...
 
## Findings
### Blocking
-...
### Nonblocking
-...
 
## Gate Evidence
-...

Then call SubmitGateVerdict:

• gate/check name: citation_qa
• status: pass, fail, or not_applicable
• evidence: artifact checks and report path.
• findings: include blocking flags where appropriate.

Pass requires concrete evidence, not just "looks good".