LawEngine built CiteBench—and the deterministic verification system that gets all 500 right. That is not a loophole. It is the point: legal citations should be checked against the source, not guessed.
LawEngine: 500/500. GPT-5.4-mini: 361/500 (72.2%). Same public pack. Same four-label task.
LawEngine deterministic verification500/500100%
GPT-5.4-mini361/50072.2%
Why this exists
This is not your grandma’s Microsoft Excel VLOOKUP.
A lookup is easy when the key and source are already clean. Legal authority is not. A one-word amendment, a bracketed exception, a line-broken PDF, an ellipsis, or a real quotation attached to the neighboring rule can change the answer. CiteBench tests whether a system can tell the difference.
Predecessor: “Each applicant shall make proof that he has completed such law study...” Current: “Each applicant shall make proof that the applicant has completed such law study...”
Exact words can be date-right and date-wrong when the applicable version changed.
Cook County Circuit Court Rule 21.07 — controlled ellipsis
“All such communications shall be ... inadmissible as evidence ... But, evidence that is otherwise admissible ... does not become inadmissible ... solely by reason of its use in a mediation session.”
A shortened quote must preserve the source order and meaning across confidentiality and admissibility.
DownloadGet the public pack and submission template.
Run locallyUse the public repository with your own system.
UploadSend predictions for aggregate scoring.
The four-label contract
Classify the citation and quote, then score the prediction file.
Verified
The quote appears in the cited authority.
VERIFIED
Not found
The quote is not in the cited authority.
NOT_FOUND
Found elsewhere
The quote is real, but belongs to another authority.
MISATTRIBUTED
Citation unresolved
The citation cannot be tied to an authority in the benchmark library.
CITATION_UNRESOLVED
Upload and score
Score a CiteBench prediction file
Download the 500-row public pack, run your system locally, and upload a CSV containing id, predicted_status. LawEngine returns aggregate accuracy, F1 scores, label counts, and a confusion matrix while the grading key remains server-side.
Methodology: CiteBench scores the public pack’s four permitted outcomes against a server-side grading key and reports aggregate results only.