Catch AI-fabricated medical citations before you publish. Verifies DOIs, PMIDs, arXiv IDs, NCT records, and metadata against public registries; grades the result in a 4-tier classification. CLI + MCP server.
Single tool 'verify_citations' has a well-structured input schema with proper types and descriptions for all 4 parameters (text, mailto, cachePath, format). Tool description is clear and domain-specific (194 chars), explaining the 4-tier classification and fabrication rate. However, the output schema is not documented in the source, the tool returns a VerifyReport object but no schema definition is visible in the MCP registration. Parameter descriptions are adequate but could be more prescriptive about constraints (e.g., format enum is declared but mailto/cachePath lack format guidance). No error handling guidance is documented. The tool follows single-responsibility principle well but lacks production-grade error recovery patterns.
Verify medical/scientific citations in a block of text against CrossRef, PubMed, and OpenAlex. Detects fabricated DOIs, invalid PMIDs, arXiv IDs, and bibliographic mismatches, returning a 4-tier classification (Verified / Bibliographic mismatch / Hallucination / Content review needed) plus a fabrication rate.
Output schema not documented. Tool returns VerifyReport (with generatedAt, totalCitations, counts, fabricationRate, citations array) but no schema definition visible in MCP tool registration. LLMs cannot plan downstream operations without knowing response structure.
No error handling guidance. Tool description does not explain what happens on network failures, invalid DOIs, or registry timeouts. LLM has no recovery path if CrossRef/PubMed/OpenAlex calls fail.
Parameter constraints under-specified. 'mailto' and 'cachePath' lack format guidance (e.g., is mailto required to be RFC 5322 compliant? Is cachePath absolute or relative?). 'text' parameter has no length limits documented, risking oversized requests.
Tool description lacks WHEN-to-use guidance. It explains WHAT the tool does but not when to call it vs. other fact-checking approaches, or what prerequisites exist (e.g., does text need to be in English? Are there citation format requirements?).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
No idempotency guarantee documented. Tool makes external API calls to registries, if called twice with identical input, will results be cached or re-fetched? Agents need to know if retries are safe.