Methodology

1. What HallucinX does

HallucinX checks every case citation in a legal brief against a full copy of CourtListener's public bulk data — 10,174,551 cases, 10,903,928 opinions and 18,130,235 citation entries from 3,361 federal and state courts, as of September 30, 2026. For each citation it reports whether the citation resolves to an opinion in that data, and whether the case name and metadata in the brief match the opinion it points to.

When the citation resolves, HallucinX compares the brief against the record: the case name, the year, and the reporter page, and it links the citation to the opinion itself. When the brief quotes the case, the comparison runs in reverse: HallucinX's server sends the opinion's text to your browser, and your browser matches the quotation against it. The brief, its text and its quotations never leave your device; only the citations do, and only as far as HallucinX's own server. A misquotation, or a quote that appears nowhere in the opinion, is flagged. Client confidentiality is preserved by design.

When the citation does not resolve, HallucinX applies a set of deterministic fabrication heuristics to what it did find — a reporter page that belongs to a different named case, a citation to a federal reporter the data covers densely that is nonetheless not found, a Westlaw number outside the plausible range for its year — and separates a citation that trips a heuristic from one the data simply does not cover. Each heuristic carries a code (H1, H2, and so on) and a reason an attorney can read to a judge; Section 4 describes each one and its known false-positive risk.

HallucinX does not evaluate the legal argument, does not check whether a case is still good law, and does not catch substantive errors in legal reasoning. Those questions turn on judgment — how a holding applies to these facts, whether a later court has narrowed it, whether the authority supports the proposition it is cited for — and that is what legal training and a license exist to do. HallucinX is built to make that time more productive: the citation check that took hours now takes minutes, with a reason recorded for every result. It is an audit aid, not a substitute for attorney review.

The verification is deterministic. Every classification has a concrete reason an attorney can show a judge. There is no language model of any kind in the verification path — no large language model (LLM), no generative AI, no statistical text model.

2. How it works

The provenance chain runs end to end as follows. An attorney walking a judge through the tool's behavior should be able to point at each step.

1. The brief is opened in your browser, locally, as a PDF.
2. Citations are extracted in your browser by HallucinX's citation parser, built on the eyecite library with HallucinX's own extraction and resolution layers on top, and reduced to the fields a lookup needs — case name, volume, reporter, page, pinpoint, court and year. Only those fields are sent.
3. The citations go to HallucinX's server and no further. The server resolves them against its own copy of CourtListener's data — the one described in Section 1 — and makes no request to CourtListener, or to anyone outside HallucinX's own hosting providers (listed in the privacy policy), during a verification.
4. For each citation, the engine looks up the matching opinion in that copy, following CourtListener's own citation-lookup rules. HallucinX then compares the case name and metadata in the brief against the record it finds. Disagreements feed the fabrication heuristics in Section 4.
5. Each citation is assigned one of five states based on what was found.

What leaves your browser, and what stays

• Stays on your device: the document, its text and its quotations, the annotated brief and the summary report. Both downloads are built in your browser.
• Sent to HallucinX: the citations, each with whether the brief quotes it and whether it sits in a table of authorities or the body, and the document's length in characters.
• Sent anywhere else: nothing from your brief. HallucinX's server makes no request to CourtListener, or to anyone outside HallucinX's own hosting providers (listed in the privacy policy), during a verification. Your browser downloads the checking code and its data packages from jsDelivr and PyPI; those hosts see your browser's address, not your document or its citations.
• Comes back: a classification and its reasons for each citation, and, for a quoted citation, the text of the opinion it cites. Your browser compares the quotation with that text.

The copy of CourtListener's data is as of a fixed date, and every report prints that date in its footer. An opinion added to CourtListener after that date is not in the copy, and a citation to it reads as not found. For a very recent citation, confirm it on CourtListener or with the issuing court.

3. The five states

Every citation in a HallucinX report carries one of five classifications. The classifications are exclusive — a citation lands in exactly one state — and ordered below by how much attention they require.

Verified — the citation exists and the metadata matches. This is the state where automatic verification has succeeded. Spot-checking is still part of the attorney's duty of competence; the other four states call for closer follow-up because automatic verification fell short or did not apply.

Caution — the citation exists but a discrepancy was detected. Common causes are case-name drift between the brief and the indexed opinion or a year mismatch. Caution is most often a real case with a clerical error in the brief; sometimes it is the early sign of a fabrication. Review before relying on the cite.

Check manually — the citation could not be verified automatically, and the cited reporter or court appears to be outside CourtListener's coverage, or the opinion postdates the copy of its data HallucinX checks against. State courts and unpublished opinions land here most often. This is not a fabrication signal. Confirm the citation by another means — Westlaw, Lexis, the issuing court, or the official reporter.

Unverified — no match was found in CourtListener's public data as of the date on the report, and no fabrication signature matched. The citation could be a real case in a coverage gap, or a fabrication that does not match a known pattern. Confirm by another means before relying on the cite.

Alert — high-confidence fabrication. One or more deterministic fabrication heuristics fired. The next section enumerates them. An Alert classification means the lookup's result is structurally inconsistent with the citation being a real case; treat as fabrication unless and until proven otherwise.

4. The fabrication heuristics

When the lookup finds no match, or finds one that disagrees with the brief, HallucinX applies a small number of deterministic heuristics to decide whether the result is a coverage gap or a fabrication. Each heuristic has a code (H1, H2, and so on) that appears in the report's reasons column for any citation it fires on. Each is the result of corpus analysis against known fabrications and known-good briefs, and each carries a known false-positive risk that an attorney reviewing a flag can weigh.

H1 — Reporter points to a different case

Pattern. The citation resolves in CourtListener's public data, but the case name on that record and the case name as it appears in the brief share zero meaningful words. “Meaningful” means longer than two letters and not a common stopword (Inc, Corp, the, of, in, and, etc.). Pincite-form short citations and string-citation extraction failures are excluded.

Why this is a fabrication signal. A real citation cannot disagree with the reporter on every meaningful word of the case name. The brief is asserting that volume, reporter, and page point at a named case; the reporter's database says they point at an entirely different one. The classic AI tell is an invented case name pasted onto a real reporter citation that belongs to an unrelated case.

Example. In Mata v. Avianca, the brief cited “Peterson v. Iran Air, 905 F. Supp. 2d 121 (D.D.C. 2013).” That reporter location does resolve in CourtListener's public data — but to a different case (an ISS Marine Services matter), with no meaningful word in common with “Peterson” or “Iran Air.”

False-positive risk. Low. Real briefs do not contain full case-name disagreements with the reporter's index.

H2 — Federal reporter citation not found

Pattern. The lookup finds no match for a citation to one of the federal reporters the data covers densely — U.S., S. Ct., F.2d, F.3d, F.4th, F. Supp. 2d, F. Supp. 3d, F.R.D. — where the cited year is more than six years old (fifteen for U.S. Reports) and the cited volume is at or below the volume the data has been measured to cover for that reporter. Government-prosecution captions (“United States v. X,” “People v. X,” “State v. X,” “Commonwealth v. X,” municipal actions) and pincite-form short cites are excluded.

Why this is a fabrication signal. CourtListener's coverage of these published federal reporters is dense for opinions more than six years old (fifteen for U.S. Reports); the most recent years have indexing lag, which is why the heuristic excludes them. A 2017 F.3d citation that is not found is unlikely to be a coverage gap: a real case at that citation would usually be indexed. That coverage is not gap-free, so confirm an H2 flag manually.

Example. In Mata, the brief cited “Varghese v. China Southern Airlines Co. Ltd., 925 F.3d 1339 (11th Cir. 2019).” The citation is not found. F.3d is one of those reporters, and 2019 is more than six years old.

False-positive risk. Low.

H3 — Unpublished state-court citation not found

Pattern. The lookup finds no match for a citation formatted as one of the recognized non-precedential state-court formats (Illinois Rule 23 orders, New York unpublished slip opinions, and similar).

Why this is a fabrication signal. Unpublished state dispositions are the formats AI fabrications most often mimic, because they are the hardest to verify by hand and are non-precedential, which reduces the chance a reader will check. CourtListener's coverage of unpublished state opinions is genuinely thin, so finding no match alone is not decisive. Combined with the format itself — which is what the heuristic gates on — it becomes a meaningful flag.

Example. In Mata, the brief cited “Shaboon v. EgyptAir Airlines Co., 2013 IL App (1st) 111279-U.” The citation matches the Illinois -U format and is not found.

False-positive risk. Medium. CourtListener does have real coverage gaps in unpublished state opinions. Treat an Alert classification backed only by H3 as a strong prompt to verify by another means rather than as a finding of fabrication on its own.

H4 — State-appellate Westlaw citation not found

Pattern. The lookup finds no match for a Westlaw citation (“YYYY WL NNNNNNN”) in state intermediate-appellate context — a parenthetical naming a state appellate court (Tex. App., Ga. Ct. App., Ill. App. Ct., App. Div., and similar) — where the cited WL year is old enough to be reliably indexed and a brief-side case name is present. California Court of Appeal citations from before 2016 are excluded: CourtListener never held that era's depublished opinions, so a citation that is not found there receives a note rather than an Alert.

Why this is a fabrication signal. Westlaw citations to state intermediate-appellate opinions more than two years old are usually indexed in CourtListener. That coverage is not gap-free, so a citation that is not found in this context, on a plausibly-formed case name, is more consistent with fabrication than with a coverage gap, and still worth confirming manually.

Example. In Mata, two such citations appeared: “Martinez v. Delta Air Lines, Inc., 2019 WL 4639462 (Tex. App. Sept. 25, 2019)” and “Estate of Durden v. KLM Royal Dutch Airlines, 2017 WL 2418825 (Ga. Ct. App. June 5, 2017).” Neither is found, both carry state-appellate parentheticals, and both are old enough to be indexed.

False-positive risk. Medium-high. State Westlaw coverage in CourtListener is the least uniform of the heuristic's inputs. Treat as a strong prompt for manual verification.

H5 — Westlaw number outside the plausible range for its year

Pattern. The Westlaw sequence number in a citation (“YYYY WL NNNNNNN”) is above the plausible maximum for that year. This heuristic fires regardless of the lookup's result — the WL number is arithmetically impossible whether or not the lookup finds a match.

Why this is a fabrication signal. Westlaw assigns sequence numbers monotonically within each year. An impossibly high number cannot have been issued; there were not that many opinions indexed that year. The number is wrong on its face.

Example. A fabricated cite of “2019 WL 11199934” carries an 11.2-million WL sequence number for 2019 — far above any sequence number Westlaw actually issued that year.

False-positive risk. Very low. The ceiling is set well above any observed real-world WL number.

Across all five heuristics, HallucinX was tuned conservatively. Gating conditions are tight, and a citation in a genuine coverage gap is treated as Check manually or Unverified rather than Alert. In benchmark testing against 774 citations drawn from a clean corpus of recent federal-circuit opinions, approximately 10.85% of legitimate citations were classified as Unverified or Alert and required attorney follow-up. Opinions are a stricter test bed than briefs — opinions string-cite and short-form-cite in ways that introduce extraction noise — so briefs, which are the tool's intended input, are projected to flag at a lower rate. On the Mata v. Avianca corpus, all six fabricated citations classify as Alert.

Separately, 72 briefs were generated with current large language models and run through HallucinX. Of 2,888 citations, 273 — 9.5% — carried a quotation that does not appear in the opinion cited, each confirmed by human review against the opinion text, in the briefs we generated. A quotation attributed to an opinion that does not contain it is an LLM hallucination.

5. Known limitations

HallucinX's confidence comes from a narrow scope. The list below is the full inventory of what the tool does not do and what it does not cover. An attorney defending the use of the tool to a judge should be able to point at this section as the boundary of the tool's claims.

Coverage limitations

• State-court coverage in CourtListener is uneven. Some jurisdictions are nearly complete; others are sparse. A “Check manually” classification is particularly common for state appellate opinions outside the coverage set.
• Unpublished opinions are inconsistently indexed across all jurisdictions.
• Recent opinions may not be in HallucinX's copy of CourtListener's data. The copy is as of a fixed date, printed on every report, and an opinion added after it reads as not found. Recent reporter volumes can also take months to reach CourtListener itself.

What the tool does not do

• HallucinX does not check whether a case is still good law. It is not a substitute for Shepardizing or KeyCite.
• It does not evaluate substantive legal reasoning. A brief can have all-verified citations and still be wrong on the law.
• It does not catch quote-misattribution where a quote does exist in the cited opinion but is being used out of context.
• A quotation that sits beside a rule's advisory committee note, a record citation or another filing is reported as a note rather than flagged, because its words usually belong to that source and not to the case; a small share of those quotations are misquotations of the cited case.
• It does not judge whether a pinpoint citation supports the proposition it is cited for. Where the cited page is available, the report shows an excerpt from it so the attorney can make that judgment.
• It does not verify that parallel citations point to the same case, though it does verify each citation independently.

What the tool may miss

HallucinX verifies citations the extractor finds. Citations the extractor fails to find pass through unverified — and silently. HallucinX's citation parser, applied to the text the browser extracts from the PDF, is robust on standard typography but can miss citations distorted by:

• OCR errors in scanned briefs. A scanned PDF whose OCR misread a digit, a comma, or a reporter abbreviation can yield a citation string the parser does not recognize.
• Unusual whitespace and line breaks. Non-breaking spaces, tabs, soft hyphens, or citations broken across line, column, or page boundaries can prevent the extractor from assembling a complete citation string.
• Non-standard typography. Smart quotes, ligatures, Unicode dashes other than the standard hyphen-minus, and unusual font substitutions can produce citation strings the parser does not match.

The practical mitigation is a citation-count check. After running a brief through the tool, count the citations in the brief by hand or by skim and compare to the count HallucinX reports. A mismatch is worth investigating: it may mean a citation was missed by the extractor and is not represented in the report at all. HallucinX does not claim a coverage guarantee on what the extractor sees; an attorney's own count is the check on what it does not.

What remains the attorney's

• Review every classification. Caution, Check manually, and Unverified require closer attention; Verified citations are not exempt from spot-checking.
• Treat the tool as an audit aid, not a final verification.
• Comply with their jurisdiction's professional responsibility rules regardless of tool output. HallucinX does not, and cannot, discharge those duties for the attorney who files the brief.

6. Why there is no language model in the verification path

The most common form of citation hallucination is a language model producing a citation string that does not correspond to a real case. Using another language model to verify the first one's output reproduces the same failure mode at a different layer. Probabilistic verification of probabilistic generation is not verification.

HallucinX uses deterministic comparison against a copy of CourtListener's public data. When the tool raises an Alert, one of the five heuristics in Section 4 fired, and the report names which. When the tool marks a citation Verified, that data holds an opinion at that reporter location whose case name and metadata match what the brief claims. Every classification has a concrete reason that can be reproduced by anyone with access to the same public data.

A judge does not have to trust the tool's judgment. The judge can verify the tool's judgment.

7. Auditability

Verification by HallucinX rests on artifacts a judge or opposing counsel can independently inspect.

• CourtListener is public. Anyone can look up any citation HallucinX classifies on CourtListener, or download the same public bulk data HallucinX checks against.
• The fabrication heuristics are deterministic and documented. Every flag carries its reason — the report names the specific check behind it and states it in plain English, so a reviewer can weigh it. There is no unexplained “the system flagged this.”
• The citations sent to HallucinX's server are not confidential client information. Under ABA Model Rule 1.6 and its state analogues, public references — citations to published judicial opinions — are not client confidences. HallucinX sends no document contents or client-identifying detail to its server, only the citations themselves, and sends nothing to CourtListener. (See Ethics and use for the bar association guidance on this point.)
• HallucinX retains no documents. Briefs are parsed in the user's browser, and the annotated brief and summary report are built there too; the original document never reaches HallucinX's servers. No log of brief content is kept. (See Privacy for details.)

8. Contact

Methodology questions from attorneys, judges, or opposing counsel can be directed to contact@hallucinx.com. We do not promise a particular response time.

Nothing on this page constitutes legal advice. Attorneys remain responsible for verifying citations and complying with their jurisdiction's professional responsibility rules regardless of the output of this or any tool. Spot-check every citation. HallucinX is built to make that workflow efficient, not to replace it.