Reference Verify Skill

Follow the procedure in .claude/rules/references.md § ハルシネーションチェック.

For each paper:

1. If DOI is available: WebFetch https://doi.org/<DOI> and confirm HTTP 200/302 resolution
2. If URL is available (no DOI): WebFetch the URL and confirm that the page contains the paper's title or author name
3. If neither resolves: WebSearch for the paper title + authors. If found, update the DOI/URL. If not found after search, mark as FAIL

Scale-appropriate execution:

• 1-5 papers: verify inline
• 6+ papers: spawn subagents to batch-verify in parallel (group by ~10 papers each)

Record each result:

• PASS: DOI/URL confirmed — paper is real
• FAIL + RE-SEARCHED: original DOI/URL failed but paper found via search — update the reference with correct DOI/URL
• REMOVED: paper not found via any method — exclude from output

Step 2: Missing Metadata Triage (optional)

When the invoking context requests metadata retrieval (e.g., populating limit fields for a literature survey), classify papers with missing metadata into three categories:

1. Paywall barrier: Papers behind publisher paywalls (IEEE, Elsevier, Springer, ACM, etc.) where full text was inaccessible. Potentially resolvable with fetch_with_auth MCP tool
2. Rendering failure: arXiv papers where ar5iv failed to render. Potentially resolvable by retrying ar5iv, alternative mirrors, or careful Semantic Scholar abstract extraction
3. No Limitations section expected: Survey/review/textbook papers that inherently lack a Limitations/Future Work section. No retrieval will help

Present these categories to the user with paper counts and titles, and ask per-category whether to attempt additional retrieval:

• Paywall: offer to use fetch_with_auth (requires valid cookies)
• Rendering failure: offer to retry with alternative access methods
• No Limitations section: inform the user these will remain as-is (no action)

Proceed only after the user responds. If declined, move on.

Retrieval sources (in priority order):

• ar5iv: https://ar5iv.labs.arxiv.org/html/<PAPER_ID> for arXiv paper full text
• fetch_with_auth: MCP tool for paywalled publisher access (IEEE, Elsevier, Springer, ACM) — uses institutional authentication cookies
• Semantic Scholar: https://api.semanticscholar.org/graph/v1/paper/<ID>?fields=abstract,tldr for TLDR or abstract hints when full text is unavailable
• Citing papers: limitations noted by papers that cite the target paper

If fetch_with_auth returns a session-expiry error, inform the user that cookie re-export is needed (see .claude/mcp/academic-fetch/README.md).

Step 3: Report

Produce a verification summary. The format depends on context:

Standalone invocation — print the report directly:

## Reference Verification Report

- Papers checked: N
- Passed: N
- Failed and re-searched: N
- Removed (unverifiable): N
  - [list titles if any]

If metadata triage was performed, append:

## Metadata Coverage

- Papers with metadata retrieved: N / M (X%)
- Papers with metadata unavailable: N, breakdown:

| Category | Count | Papers | Action taken |
|----------|-------|--------|-------------|
| Paywall barrier | N | [list keys] | [action] |
| Rendering failure | N | [list keys] | [action] |
| No Limitations section | N | [list keys] | N/A |

Called from literature-survey — return structured data so the survey can embed the results into its Survey Methodology section. The survey skill specifies the exact output format it expects.

Usage from Other Skills

Other skills can invoke this skill by including an instruction like:

Use the reference-verify skill to verify all papers before adding them to MAIN.md.

The calling skill should specify:

• The list of papers (or point to the file containing them)
• Whether metadata triage is needed (and which fields)
• The desired output format (inline report vs. structured data for embedding)