LexRails

Deterministic legal-citation fabrication checks for AI agents.

USDC on Base · x402no keys · no signupEd25519-signedoffline · deterministicno LLM in the path

Language models fabricate case citations — 700+ documented court incidents, real sanctions. A large class of fabrications is catchable structurally, without any database: impossible reporter volumes, unknown reporters, malformed statutes. LexRails sells exactly that check, per call, to the agent drafting the document.

GET /v1/cite/check?q=410+U.S.+113
→ "plausible" — well-formed, within U.S. Reports volume range

GET /v1/cite/check?q=999+U.S.+999
→ "impossible" — U.S. Reports has ~610 volumes; strong fabrication signal

Endpoints

`GET /v1/cite/check?q=` One citation, structurally checked: impossible / unknown-reporter / plausible. Case reporters + Westlaw/LEXIS vendor formats.	$0.008
`POST /v1/cite/extract` Send a whole draft (`{"text": …}`): extracts every citation, checks each, returns a fabrication riskFlag. The pre-filing gate for drafting agents.	$0.028
`GET /v1/uscode/validate?title=&section=` U.S. Code structure: titles 1–54 (53 reserved), section format.	$0.009
`GET /v1/cfr/validate?title=&part=&section=` CFR structure: titles 1–50 (35 reserved), 10b-5-style sections.	$0.009
`GET /v1/ecli/validate?id=` European ECLI: format, jurisdiction, year.	$0.009
`GET /v1/reporters` The bundled reporter table with volume bounds.	free

Honest scope: verdicts are structural. "Plausible" is not existence proof — for existence, query a case-law database (e.g. CourtListener). LexRails catches what's provably fake, offline, deterministically, and signs every answer (Ed25519) so it can be verified against /.well-known/signing-key.json.

For agents

llms.txt · openapi.json · x402 discovery