---
license: other
license_name: mixed
pretty_name: Citation-Checker Benchmark
language:
- en
size_categories:
- n<1K
task_categories:
- text-classification
tags:
- legal
- citations
- hallucination-detection
- benchmark
configs:
- config_name: charlotin-train
  data_files:
  - split: train
    path: data/charlotin-train/labels.csv
- config_name: charlotin-test
  data_files:
  - split: test
    path: data/charlotin-test/labels.csv
- config_name: ai-gen
  data_files:
  - split: test
    path: data/ai-gen/labels.csv
---

# Citation-Checker Benchmark

A benchmark of whole legal briefs with hand-labelled citations, together with the reports that
four citation checkers produced on those briefs: Westlaw, Lexis+, isthiscasereal.com and
Hallucite. Every checker was given the same document and its report was aligned to the labels.

The inputs and labels are in `data/`; the checkers' raw output and our parsed results are in
`results/`.

## Components

Each component is a separate config. Components are scored and reported separately, never
pooled.

| config | split | documents | labelled citations | what it is |
|---|---|---|---|---|
| `charlotin-train` | `train` | 40 | 567 | Filings from cases in Damien Charlotin's AI Hallucination Cases database, mostly cases dated March-July 2025, plus a few early well-known cases and every S.D. Tex., W.D. Okla. and E.D.N.Y. case in the database at the time. Compiled in July 2025. |
| `charlotin-test` | `test` | 9 | 158 | A random sample of cases in the same database, taken in October 2026. |
| `ai-gen` | `test` | 16 | 214 | Motions to dismiss drafted by LLMs from real federal complaints. |

Together that is 65 documents and 939 citations.

**`charlotin-train` was used to improve Hallucite**, so it is not a fair benchmark for comparing
Hallucite with the other checkers. The split names reflect this. `charlotin-test` and `ai-gen`
played no role in developing Hallucite.

### The Charlotin components

Charlotin's database typically links only to a court order, not to the filing that contained the
hallucinated citations. For each case we tried to find the original filing. When we could not
obtain it (filings were sometimes stricken from the record), we either skipped the case or used
the court order instead, so some documents here are orders rather than filings. The database
stores metadata about the cases, not the documents; the text is public court record, included
here as extracted.

### `ai-gen`

Complaints filed in 2024 or 2025, with embedded text between 10,000 and 32,000 characters, were
sampled evenly across S.D. Tex., W.D. Okla. and N.D. Okla. Each of eight models was given a
complaint and a minimal prompt and asked for a motion to dismiss: GPT 4o, GPT 5, GPT 6-Sol,
Claude 4 Sonnet, Claude 5 Sonnet, Gemini 2.5 Flash, Gemini 3 Flash and Gemini 3.8 Flash. The
models were accessed through OpenRouter with default temperature, default reasoning levels and
no tools (no internet access). No model was asked to fabricate citations. The briefs are model
output, not real filings.

## Files

```
data/<config>/<document>.txt        the document's text (also .docx: the same text, reflowed)
data/<config>/labels.csv            one row per citation occurrence (the dataset viewer shows this)
data/<config>/documents.csv         per-document metadata (briefs.csv for ai-gen)
results/<system>/<config>/          each system's raw output, exactly as produced
results/<system>/ledger.<config>.csv   our parsed result: one row per labelled citation
```

A document's id is its filename stem. Raw results and the label file's `case_file` column use
the same stem, so everything joins back to its input.

The `.docx` files were uploaded to Westlaw and Lexis; they carry the same text as the `.txt`
with lines reflowed into paragraphs and page-number lines dropped, because Lexis cannot read a
citation that crosses a line break. isthiscasereal.com was given the same reflowed text (pasted,
in pieces of at most 15 citations, since it accepts at most 20 per request).

## Labels

For each document we extracted all full case citations (short-form citations and citations
without a case name are excluded) and labelled each by hand, based on a search of Westlaw or
Lexis. `labels.csv` has one row per citation occurrence.

| column | meaning |
|---|---|
| `case_file` | the document's filename stem |
| `cite_idx` | 0-based position of the citation in the document, in order of appearance |
| `citation_text` | the citation as written |
| `citation` | the cite alone, `volume reporter page`, normalized (e.g. `556 U.S. 662`) |
| `validity` | `strict`, `referential` or `invalid` |
| `parallel_citation` | a second cite to the same case, where the brief gave one |
| `correct_case_name`, `correct_citation` | for `referential` rows, the correction |
| `src_same_case_name`, `src_same_citation` | for an `invalid` citation that resolves to a real case with a different name, that case's name and citation |
| `src_notes` | provenance notes carried over from the source |

`validity`:

- **`strict`**: the citation is valid, allowing for reasonable variations in the case name.
- **`referential`**: a reader can recover the intended authority despite an understandable typo
  in one party name or one component of the reporter citation, with no other errors.
- **`invalid`**: the authority does not exist, or the reader cannot recover it from what is
  written. An invalid citation whose reporter cite points to a real case with an entirely
  different name is a *collision* (`src_same_citation` is filled in); otherwise it is
  *nonexistent*.

Citations are judged as the author wrote them: OCR or extraction noise in the text is not the
author's error. Systems are scored under two standards: flagging only `invalid`, and flagging
`invalid` plus `referential`.

Non-case authorities such as statutes and rules are not labelled.

## Systems and results

| system | product | input | run dates (2026) |
|---|---|---|---|
| `westlaw` | Westlaw Advantage, Litigation Document Analyzer ("Check your work") | the `.docx` | Oct 1-9 (`charlotin-train`), Oct 3-4 (`charlotin-test`), Oct 6-7 (`ai-gen`) |
| `lexis` | Lexis+ Brief Analysis | the `.docx` | same dates as Westlaw |
| `isthiscasereal` | isthiscasereal.com | the reflowed text, in pieces | Oct 8-14 (`charlotin-train`), Oct 8 (others) |
| `hallucite` | Hallucite | the `.txt` | Oct 14 (`charlotin-train`), Oct 4 (`charlotin-test`), Oct 7 (`ai-gen`) |

These products change without notice, so the dates matter. Raw output is stored as the
system produced it: Westlaw's full report as `.rtf` (Word download), Lexis's report page as
`.mhtml` (Chrome "Single File"), isthiscasereal.com's report pages as `.html`. Hallucite has no
raw folder here, only a ledger.

### Ledgers

`results/<system>/ledger.<config>.csv` is our parse of the raw output, joined to the labels. It
can be regenerated from the raw files and the labels. One row per labelled citation.

| column | meaning |
|---|---|
| `system`, `component`, `case_file`, `cite_idx`, `citation`, `citation_text` | which citation |
| `finding` | `citation_issue` (the system explicitly warned), `listed` (returned, nothing raised) or `absent` (the output never mentions it) |
| `flagged` | 1 if `finding` is `citation_issue` |
| `found` | 1 if the system mentioned the citation at all |
| `validity` | the label, copied from `labels.csv` |
| `vendor_case_name`, `vendor_citation` | what the system printed for this citation |
| `joined_on` | how the system's entry was matched to the label: `citation` (exact), `parallel_citation`, `case_at_citation`, or `pincite` |
| `vendor_section`, `vendor_resolution`, `vendor_reason`, `treatment_flag` | system-specific fields |

How findings are read:

- A citation is invalid if the system flags it as problematic, and valid otherwise. A citation
  the system's output never mentions is `absent` and is counted as wrong whatever the label:
  silence is not the same as checking a citation and finding nothing.
- No case-name matching is involved. If a system lists a different case under the citation and
  says nothing, that is its judgment (`listed`); `vendor_case_name` beside `citation_text` lets
  you see the substitution.
- Treatment flags (KeyCite, Shepard's) are not read as validity signals.
- Hallucite's `invalid`, `uncovered` and `unpublished` statuses are warnings; `exists` and
  `incomplete` are `listed`. `uncovered` means the reporter is outside Hallucite's data, so it
  could not confirm the citation; scoring that as a flag matches how the other systems treat
  citations they hold no data for.

## Limitations

- Labels are hand-made and may contain errors.
- `charlotin-train` was used to improve Hallucite (above).
- Westlaw, Lexis and isthiscasereal.com were used through their web interfaces and Hallucite
  was run locally. Vendor products change, so these are results for those dates, not for the
  products in general.
- Some Charlotin documents are court orders rather than the original filings.
- `ai-gen` briefs contain fabricated citations produced by language models. They are not
  real filings.

## License and provenance

The labels, ledgers and this README are released under CC-BY-4.0. The Charlotin documents are
public court records. The `ai-gen` documents are text generated by language models from public
complaints. The files under `results/westlaw`, `results/lexis` and `results/isthiscasereal` are
the output of those services and remain subject to their owners' terms.

To request removal of anything here, contact bl759@cornell.edu.