A quarter of the citations in filed briefs are Westlaw or Lexis ids. We list them, we do not guess.Coverage ›
PROOFREAD.LAW
Proofread

What is checked,
and how often it is wrong

A checker that hides its blind spot is the thing it is supposed to catch. Every number below was measured on filed briefs, court decisions and machine-drafted paragraphs. Each one, with its sample size, date, engine version and caveats, is on Measurements.

89%

of invented cases caught (187 of 209 that courts found and that carry a reporter citation)

14 / 15

invented quotations caught

1.2%

false flags on the court's own writing

No language model in the default check

It is register lookup, name matching and quotation matching.

The register

sourcewhat it holdsas of
CourtListener bulk data10.1 million cases, 18.1 million citations, one opinion text per caserefreshed daily from the search API
Caselaw Access Projectopinion text where CourtListener has none (to 2019)static
govinfofederal appellate opinions, all 13 circuits, 2018 onward (138,881 opinions; 28,700 cases the bulk data lacked)daily
Supreme Court DatabaseU.S. Reports citations for recent terms the bulk data lacks2025 release

Blind spots, stated

  • Westlaw and Lexis identifiers (2023 WL 1234567) cannot be resolved by any open register. A quarter of the citations in filed briefs. Listed, never flagged.
  • The bulk data lacks 10 to 15% of federal appellate opinions from 2019 to 2023. The register carries the published opinions of all 13 circuits from 2018 onward from govinfo, matched 95 to 100% for every circuit except the 2d (87%). A miss in a matched circuit is a finding; in the 2d Circuit, and for unpublished opinions everywhere, it is shown as "cannot verify".
  • Statutes, regulations, rules, treatises and law reviews are not checked.
  • Whether a later court has overruled or limited a case, or a higher court reversed or vacated it on appeal ("is it still good law"), is not checked. This is not a citator; that is what KeyCite and Shepard's are for. Every report says so in its coverage statement.

Switzerland

A draft that cites Swiss law is checked against the Swiss register: BGE/ATF/DTF volumes, Federal Supreme Court dockets (4A_281/2015), the federal courts and the courts of the 26 cantons; German, French and Italian. The report says which register it used.

courtdecisions heldfromshare of the live entscheidsuche.ch index
Bundesgericht, published (BGE)50,5031875100%
Bundesgericht, all decisions (BGer dockets)193,3892000 (a few from 1986)99.95%
Bundesverwaltungsgericht108,991200799.1%
Bundesstrafgericht12,078200497.7%
Bundespatentgericht204201296.5%
The 26 cantons749,434 (VD 211,099 · GE 174,357 · ZH 100,445 · TI 67,085 · BL 22,458 · BE 22,031 · SG 20,299 …)varies by cantonnot measured per canton

1,147,178 decisions in all; the live index checked daily.

  • The Federal Supreme Court's published decisions (BGE) are held from 1875; its unpublished decisions are online only from 2000, as the court itself publishes them (complete from 2007). An unpublished docket from before 2000 is shown as "cannot verify", never as a finding.
  • A BGE citation whose volume the register holds but whose page no decision occupies is a finding; a volume newer than the register is "cannot verify".
  • 13,675 decisions carry fewer than 500 characters of text (cantonal abstracts; Obwalden's 2,200 decisions, whose server serves empty pages). The deep check cannot read those.
  • Cantonal coverage varies by canton and is not measured against a live index; a cantonal miss is "cannot verify".
  • No citator here either: whether a decision was set aside on appeal or abandoned by a change in case law is not checked.

Controls: 1,700 invented Swiss citations, all reported not found; BGE citations found 99.9% of the time with their pincite, Federal Supreme Court dockets 96.2%. The deep check, on 757 held-out claim-and-decision pairs, at the confidence the report shows, confirms 73.2% of supported claims and marks a wrong passage as support for 2.2% of wrong pairs (jev 1.14, 2026-10-30; 76.5% and 3.1% on jev 1.13). That set has been run more than once, so these numbers are not blind (measurements). Not yet measured on filed Swiss briefs; that measurement is next.

How the numbers were measured

528 citations from court decisions that sanctioned invented or misused authorities; 7,067 citations in 141 briefs filed in 29 federal district courts in 2024 and later; 300 paragraphs drafted by three language models against the court's own 100 paragraphs as the floor. In the filed briefs, no citation was to a case that does not exist; in the machine drafts, 7 to 21% of citations pointed to no case or to a different case.

The deep check (opt-in), at the confidence the report shows, confirms 55.0% of good citations and marks a wrong passage as support for 4.5% of wrong pairs; at any confidence, 67.7% and 11.3%. That is why it is a review queue with the passage, not a verdict. Every measurement, with its caveats.