What it does
verify_citations checks legal citations against the corpus. Pass a list of citations, a drafted answer (its citations are extracted and an uncited-claims scan runs), or both. Each citation comes back with whether it exists, the document it resolves to, its temporal status, and its provenance — whether the calling AI was actually shown that document in this conversation. It merges the earlier single-citation and whole-answer checks into one tool.
Inputs
| Name | Type | Required | Description |
|---|---|---|---|
citations | string[] or object[] | One of | Citations to check: ECLI, docket, act number, § 63 zákona č. 311/2001 Z. z., slov-lex or EUR-Lex URL. |
answer | string | One of | A drafted answer; its citations are extracted and claims without a citation are flagged. |
receipts | string[] | No | Receipts from earlier calls in this conversation; they prove which documents were retrieved. |
Returns
For each citation:
existsand the resolved document (id, title,source_url), ornot_found.- temporal status of a legislation citation — in force, not yet in force, superseded by a newer version, repealed, or unknown where the record has no effective dates. Case-law citations have no temporal status.
provenance_status—returned(a receipt proves the calling AI was shown it),exists_not_returned(real, but nothing shows it was retrieved in this conversation), ornot_found.- For a docket shared by several decisions, the candidates, so the AI disambiguates instead of the tool guessing.
For an answer, the list of claims that carry no citation (uncited_legal_claims) and uncited_claims_scan, which says whether the scan ran (not_requested when no answer was passed, so an empty list then means nothing was scanned). receipt_checks gives one verdict per receipt you passed.
Example
The calling AI has drafted a reply citing three decisions and two provisions. Before sending it, it calls verify_citations with answer set to the draft and receipts from its earlier lookups. Two citations come back returned, one exists_not_returned (the AI remembered it rather than looking it up), and one not_found. It opens the second with Read a decision and drops the last.
Notes
- Deterministic, no LLM involved.
- Debits the corpus quota (see Usage limits).
- Found is not the same as current, and current is not the same as retrieved: the three fields are reported separately so none is inferred from another.