Real filings. Real sanctions. Avoidable ones.

Mata is the origin story. Steven Schwartz of Levidow, Levidow & Oberman used ChatGPT to research an aviation-injury opposition and filed an affirmation citing Varghese, Shaboon, Petersen, Martinez, Durden, and Miller, none of which exist. When Judge Castel ordered the actual opinions produced, the response was a set of fabricated excerpts ChatGPT also generated. The $5,000 Rule 11 sanction was small; the reputational cost was not, and the case launched the wave of standing orders requiring AI-use certification that defines the category today.

Run on the same brief, Sentinel Citation verifies every citation against CourtListener, Cornell LII, and eCFR before any judging begins. The six fabricated cases return no authority to fetch, so accuracy, relevance, and treatment collapse to zero and the brief bottoms out at 22/100. Existence verification is the floor, not the ceiling, but it is exactly the floor Mata fell through. Run before filing, that one report turns a career-defining sanction into a five-minute fix.

Read the executive report

Offending brief, ECF 21

CourtListener docket

Coverage

Gauthier is the case that breaks the comforting assumption that a second tool catches the first tool's mistakes. Counsel drafted the response with Claude, then ran it through an unnamed AI cite-checker for safety, and still filed two nonexistent cases and fabricated quotations supporting the brief's core Sabine Pilot wrongful-termination theory. The court was unimpressed by the verification step and imposed $2,000 plus a mandatory AI-ethics CLE.

This is where Sentinel Citation's real value shows, and where we are precise about its edges. Sentinel Citation is less a hallucination alarm than a strength evaluator: it validates each citation against the actual opinion and judges whether the authority supports the use the brief made of it. On Gauthier's most attackable cluster, the three Roca v. King's Creek Plantation citations, the jury scored them at the floor and flagged them as misused, reading the reporter as a real but distinguishable opinion rather than naming it an outright fabrication. What it told counsel unambiguously is that the cite would not bear the weight placed on it, that Roca is unpublished and non-precedential, and that controlling Texas Supreme Court authority (Safeshred) should carry the point instead.

Read the executive report

Offending brief, ECF 41

CourtListener docket

Coverage

Lacey is the answer to "my research subscription already does this." A partner at a firm with K&L Gates on the brief assembled the argument using three premium, name-brand tools, CoCounsel, Westlaw Precision, and Google Gemini, and still produced a filing where nine of 27 citations were defective and at least two cases did not exist. The defective brief first went to a Special Master; the sanctions order (ECF 119) reproduces the bad citations as an appendix. The $31,100 joint sanction and the court's "collective debacle" line made it the AmLaw-100 cautionary tale.

Sentinel Citation's report explains why the premium stack missed it. The most attackable citations are not invented cases; they are real cases cited backwards. Costco Wholesale Corp. v. Superior Court is characterized as permitting in camera review when Costco squarely prohibits it; Aetna is cited for a sword-and-shield proposition closer to dictum and narrowed by later authority. Existence-aware research platforms confirm those cases are real, which is exactly why they pass them. The question is not whether the case exists; it is whether the holding supports the use, and that is the layer Sentinel Citation adds. Scored before filing, the brief returns 28/100 with the privilege cluster flagged, and a $31,100 published debacle becomes a redline.

Read the executive report

Offending brief, ECF 119 appendix (pp. 20-31)

CourtListener docket

Coverage

Johnson v. Dunn is where the stakes moved past money. A large, well-resourced firm, Butler Snow, filed discovery motions carrying ChatGPT-fabricated citations, and the court concluded that fines and embarrassment had stopped deterring anyone. It disqualified three attorneys from the case, issued a public reprimand, and referred them to the bar. The message to the profession was that a line-item sanction is no longer the worst-case outcome.

This is the gallery's clearest demonstration that Sentinel Citation is a planning partner, not a pass/fail buzzer. The brief scores 55/100, the highest here, because most of its authority is sound: 26 of 48 citations come back solid. The nine vulnerable cites are not scattered filler; they cluster on the brief's load-bearing privilege and discovery arguments. Jane Student 1 v. Williams is cited with a pincite that overshoots where the opinion ends; Jaffe v. Redmond is stretched past its psychotherapist-patient holding; out-of-circuit, unpublished orders are leaned on as if binding. Sentinel Citation returns not a verdict but a map: for the filer, exactly which nine citations to fix or pull before filing; for the opponent, a ready-made targeting plan for the response.

Read the executive report

Offending brief, ECF 182

CourtListener docket

Coverage

Coomer v. Lindell is a high-volume, high-visibility example with a sequel. Counsel filed an opposition brief carrying roughly 30 defective citations, first denied using AI, then acknowledged running the brief through AI. The court imposed $3,000 on each attorney. The instructive part came later: in May 2026 a second sanction, $5,000, landed for a repeat miscitation in post-trial briefing. One clean pass had not become a habit.

The report shows both halves of what Sentinel Citation is for. As defense, a pre-filing run flags 42 of 90 citations as vulnerable, clustered on the brief's Rule 403 and 404(b) evidentiary arguments, and a re-run before the post-trial brief is precisely the gate that catches a repeat miscitation before it draws a second sanction. As offense, that same 32/100 report is a response brief's worth of ammunition: opposing counsel receives a per-citation map of which authorities are mischaracterized or unsupported, each paired with the passage that contradicts the brief's use of it. Every citation in it came from a public filing, so a skeptic can watch it run without uploading anything of their own.

Read the executive report

Offending brief, ECF 283

CourtListener docket

Coverage

Reports reflect our own published testing on these specific filings, run unmodified. Each citation is verified against CourtListener, Cornell LII, eCFR, or Scholar, then judged by a three-model jury (Claude Sonnet, Gemini Flash, GPT, versions pinned per run) and reconciled by a meta-judge into one verdict per citation. Absence from Sentinel Citation's sources is not absence from law; a not-found result is reported with its reason, not silently passed or dropped. The multi-model jury and meta-judge methodology is patent-pending (US 64/056,937).

Reports were generated from documents in the public court record. Sentinel Citation analyzes filings as submitted; verdicts reflect the sources available at run time. Nothing here is legal advice or a comment on any attorney's professional conduct beyond what the cited court orders state.