Audit Bibliography
auditBibliographyVerify a WHOLE bibliography in one call — the batch counterpart to verifyCitation. Each entry runs the same fabrication check (real, resolvable identifier paired with a title that does NOT match the resolved paper; Topaz et al., Lancet 2026) plus a retraction lookup, and the tool returns a per-entry verdict table and a corpus summary. Use when the user pastes a reference list, a .bib / .ris file, or asks to 'check all these citations at once' / 'audit my bibliography' / 'which of these references are fake or retracted'. Input: EITHER bibliography (raw BibTeX / RIS / CSL-JSON text — format auto-detected) OR claims (an array of pre-parsed {title + identifier} objects), not both. Capped at 25 entries per call; excess is dropped and reported via truncated. checks defaults to ['retraction'] (pass [] to skip); screenWithLlm opt-in per entry (same auth gating as verifyCitation). Returns: { format, entries: [{ index, sourceKey?, status: 'ok'|'error', verdict: 'matched' | 'mismatch' | 'not_found' | 'ambiguous', confidence, matched, mismatches, retraction: { checked, doi, isRetracted, hasCorrections, hasConcern, notices } | null, provenance }], parseErrors: [{ index, error, message }], truncated, summary: { total, matched, mismatch, ambiguous, not_found, not_checked, errored, retracted } }. not_checked counts entries declined before a verdict (no identifier, container-only identifier, too sparse) — report them as unverified, never as suspect. Reading the result: index is 1-BASED (entry 1 is the first reference) — do not add 1 again when reporting it. sourceKey is the entry's own key in the source file (BibTeX cite key, RIS ID, CSL-JSON id) and is the reliable way to point a user at the offending reference; it is absent on the claims[] path. entries and parseErrors share one index space, so a given input position appears in exactly one of them — report parseErrors as UNCHECKED, never as clean. summary.total counts verifiable entries only, excluding parseErrors and anything past the cap; summary.retracted is a separate axis from the verdict counts (an entry can be both matched and retracted), so never sum those fields. A non-zero truncated means the audit is incomplete — split the bibliography and call again. Per-entry leniency: one entry that fails to resolve becomes status:'error' without failing the batch. This audits citation IDENTITY (does each identifier resolve to the claimed work, and is it retracted) — it does NOT check whether a source supports the claim it is cited for. Read-only and idempotent. Works anonymously for the non-LLM path; SCHOLAR_API_KEY (a free ssk key from https://scholar-sidekick.com/account) or a paid RapidAPI tier raises rate limits and enables the optional LLM screen.
Tool call template
This tool declares no required parameters. Add optional parameters from the input schema as needed.
{
"id": 1,
"jsonrpc": "2.0",
"method": "tools/call",
"params": {
"arguments": {},
"name": "auditBibliography"
}
}Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| checks | No | Per-entry enrichment checks. Defaults to ['retraction'] (flags retracted / corrected / expression-of-concern works via Crossref + Retraction Watch, keyed on each resolved DOI). Pass [] to skip the retraction lookup. | |
| claims | No | Pre-parsed citations to audit — an alternative to `bibliography` for agents that already hold structured references. Each needs a `title` plus whatever identifiers the citation carries. | |
| format | No | Override format auto-detection for `bibliography`. | |
| bibliography | No | Raw bibliography text to parse and audit — BibTeX, RIS, CSL-JSON, or a plain-text reference list (one reference per line, numbered, or blank-line separated). Provide EITHER this or `claims`, not both. Format is auto-detected from the content; override with `format`. Plain text is split into one entry per reference and each is checked against the record its identifier resolves to; a reference without an identifier is reported as not checked. Capped at 25 entries per call (excess is dropped and reported via `truncated`). | |
| screenWithLlm | No | Opt-in Stage 3 LLM screen applied per entry (same gating as verifyCitation: authenticated first-party key or paid RapidAPI tier). Default false. | |
| semanticTitle | No | Opt-in semantic title check applied per entry (any tier; see verifyCitation). Default false. |
Output Schema
| Name | Required | Description | Default |
|---|---|---|---|
| format | No | Detected input format ('bibtex' | 'ris' | 'csl-json'), or null for the claims[] path and for plain-text input. | |
| entries | No | One result per verifiable entry: { index, sourceKey?, status: 'ok'|'error', verdict, confidence, matched, mismatches, retraction, _provenance }. `index` is 1-BASED and counts position in the submitted input, so entry 1 is the first reference — do not add 1 again when you report it to a user. `sourceKey` is the entry's own key in the source file (BibTeX cite key, RIS `ID`, or CSL-JSON `id`) and is the reliable way to map a verdict back to the user's bibliography; it is absent on the claims[] path and on formats that carry no key. `status:'error'` means that one entry failed to verify, not that the batch failed. Entries appear in input order. | |
| summary | No | Corpus roll-up: { total, matched, mismatch, ambiguous, not_found, not_checked, errored, retracted }. `total` counts verifiable entries only, so it excludes `parseErrors` and anything past the cap. `not_found` = looked up and nothing exists; `not_checked` = declined before a verdict (no identifier in the reference, a container-only identifier, or too sparse to check) — tell the user those were NOT verified rather than treating them as suspect. `retracted` is a separate axis from the verdict counts, and an entry can be both matched and retracted — never add the fields together. | |
| truncated | No | Count of entries dropped beyond the 25-entry cap. Non-zero means the audit is incomplete — split the bibliography and call again for the rest. | |
| parseErrors | No | Entries that could not be parsed or lacked a title: { index, error, message }. Same 1-based `index` space as `entries`, so the two arrays never collide: a given input position appears in exactly one of them. Report these to the user as unchecked, NOT as clean. |