A one-page reference for attorneys answering the bench's question: what verification did you perform on AI-generated citations before filing this document?
Three obligations converge on every AI-assisted filing:
In June 2023, Mata v. Avianca, 22-CV-1461 (S.D.N.Y.), sanctioned two attorneys for filing a brief that cited six entirely fabricated cases — the cases sounded real, the holdings were plausible, the citation format was correct, and not one of them existed. Subsequent reported sanctions have followed in at least eight federal districts and a growing number of state courts.
BenchSlap is the only legal-AI platform whose architecture cannot emit a citation that doesn't exist in a verified closed corpus. The relevant invention is called AEGIS:
✓ VERIFIED / ⚠ UNVERIFIED / ✗ FABRICATION.When Judge X asks counsel "What verification did you perform on the citations in this brief?", the BenchSlap-using attorney can answer with concrete specifics rather than ambiguity:
| The question | The defensible answer |
|---|---|
| Did you verify every citation? | Yes. Every citation in this brief was checked against a 3.2-million-opinion closed corpus before the brief was finalized. |
| How is the corpus authenticated? | Every authority is hash-pinned (SHA-256) at ingest time. The hash is recomputed on each verify pass; a mismatch hard-blocks the filing. |
| How do you know the AI didn't fabricate a citation? | The AI's output passes through the AEGIS gate. A citation that doesn't exist in the closed corpus cannot reach the printed brief. |
| How do you know the case wasn't overruled? | AEGIS PRIME maintains a treatment graph. When a cited case has a non-null superseded_by_opinion_id, the brief surfaces a warning and the AI rewrites with the controlling authority. |
| What if the AI quoted the dissent as the majority? | AEGIS PRIME's attribution check matches every quote to its panel role (majority / concurrence / dissent). A mis-attributed quote hard-blocks. |
| Can you produce the verification record? | Yes. Every Drafter and Auditor output exports with an embedded verification certificate listing each citation, its verdict, source tier (T0–T4), and the AEGIS hash. |
| Badge | Meaning | Action required |
|---|---|---|
| ✓ VERIFIED | Citation found in opinion library with content-hash match. Source tier T0 (cache) through T4 (multi-AI peer review). | None — cite as written. |
| ⚠ UNVERIFIED | Citation not found in the closed corpus and could not be authenticated by remote source. | Verify independently before filing. Do not rely on the AI's word. |
| ✗ FABRICATION | Citation does not exist as written. Structural fabrication (impossible year, volume, page) or AEGIS-confirmed absence from the corpus. | Remove from filing. The AI invented this citation. |
| ⚠ OVERRULED | Citation exists but has been superseded by a later opinion. Treatment graph identifies the overruling case. | Cite the overruling authority or note the cited case is no longer good law. |
The reason generate-then-verify pipelines (ChatGPT for legal work, plus other consumer LLMs marketed for legal use) cannot offer this defense: their architecture invites hallucination by design. They generate text first, then optionally ask a second AI to verify it. The second AI suffers from the same statistical failure mode as the first.
BenchSlap inverts the architecture. The corpus is closed and hash-pinned at ingest time, before any user request. The AI's role is constrained to selecting from the verified corpus — not to inventing citations and hoping a verifier catches it.
benchslap.pro/whitepapers/00–09github.com/Benchslap/Benchslap — the verification kernel (lib/authority-hash.js, lib/verification-pipeline.js, lib/post-stream-gate.js, lib/case-fact-extractor.js) is auditablebenchslap.pro/api/demo/cite-check — any citation in a filing can be re-verified independently