Return MCP configuration and dependency status.

Call this first when setting up CiteGuard or diagnosing surprising results. By default it does not contact OpenAlex/Crossref/arXiv/Semantic Scholar and does not load model weights. Set check_sources=true to run a lightweight per-source live probe using health_query.

Verify ONE citation against live scholarly sources (OpenAlex/Crossref/arXiv).

Provide either a free-text citation in raw_text, or structured fields (title, authors, year, doi, arxiv_id, venue). Returns a verdict (verified | metadata_mismatch | not_found | ambiguous), the canonical record, per-field diffs, a suggested corrected citation when confident, and which sources were checked. A not_found verdict means "could not be verified", not a definitive proof of fabrication. Evidence and full-text inputs belong to the claim-support tools, not to existence/metadata verification.

Verify MANY citations at once.

citations is a list of objects, each with any of: raw_text, title, authors, year, venue, doi, arxiv_id. Returns a per-citation report plus summary, risk_ranking, and review_summary. Agents should branch on review_summary.triage_plan, compact risk_reason, and suggested_fix.kind / suggested_fix.requires_user_confirmation instead of parsing prose. Set review_summary.suggested_fix_summary.auto_apply_allowed=false means batch repairs must be proposed to the user, not silently applied. Set high_risk_only=true to return only high-risk result rows while preserving full-batch review_summary counts and filtered.returned_indexes / filtered.omitted_indexes for original-input traceability.

Judge whether a cited paper SUPPORTS a claim sentence (abstract-level).

Resolves the paper (existence), then assesses support with a reranker+NLI ensemble. Verdicts: supported | weakly_supported | insufficient_evidence | contradicted. insufficient_evidence means the abstract does not address the claim - NOT that the paper is unsupportive. Agents may pass lawful caller-provided full_text excerpts or full_text_file paths; those are tagged as evidence_scope=full_text, and CiteGuard will not fetch gated full text. Deep models are downloaded on first use; without them the engine falls back to "heuristic" (no supported/contradicted verdicts) and says so. Set CITEGUARD_RERANKER_MODEL / CITEGUARD_NLI_MODEL to use multilingual models for non-English claims.

Judge whether a SET of cited papers supports one claim sentence.

citations is a list of citation objects accepted by verify_citation_tool. Citation objects may include lawful caller-provided full_text excerpts or full_text_file paths; those are tagged as evidence_scope=full_text, and CiteGuard will not fetch gated full text. The result aggregates per-citation support checks and exposes support_mode_details with the policy no_unstated_multi_hop_or_full_text_support. A supported aggregate requires at least one resolved citation to strongly support the claim; contradictions are surfaced as high-risk.

Search for scholarly records that may contain counter-evidence.

This tool returns review candidates only. It does not prove contradiction and does not change a support verdict; run claim-support checks on any promising candidate before editing text or replacing citations.

Judge MANY claim-citation support pairs at once.

items is a list of objects with a required claim plus any citation fields accepted by verify_citation_tool: raw_text, title, authors, year, venue, doi, arxiv_id, full_text, full_text_file, and optional per-item lang. An item may alternatively provide citations, a non-empty list of citation objects, to assess whether a claim is supported by the cited set. Returns a per-item support report plus summary, risk_ranking, and review_summary. Agents should branch on review_summary.triage_plan, compact risk_reason, and suggested_fix.kind / suggested_fix.requires_user_confirmation instead of parsing support prose or silently editing citations. review_summary.suggested_fix_summary.auto_apply_allowed=false means batch repairs must be proposed to the user, not silently applied. Items may include lawful caller-provided full_text excerpts or full_text_file paths; those are tagged as evidence_scope=full_text, and CiteGuard will not fetch gated full text. Set high_risk_only=true to return only high-risk rows while preserving full-batch review_summary counts and filtered.returned_indexes / filtered.omitted_indexes for original-input traceability.