THE REVIEW SCORING BENCH
Score an AI contract review.
Every citation checked against the contract.
Paste the review your tool produced for one of the synthetic contracts.
The harness scores clauses, risks and citations against a hand-authored answer set and applies the release policy.
- 01 Choose a contract
- 02 Paste the review
- 03 Score & export
Review under test
Loading the answer set…
Your review stays in this tab and is cleared when you close it. The Python runtime downloads on first use. The review must follow the review schema; citations must quote the contract verbatim.
SOURCEThe contract
Synthetic. Citations are matched against this text after normalising case, punctuation and whitespace.
Scorecard
Your scorecard starts here
Load an example or paste your tool's review JSON,
then score it against the answer set.
Deterministic code. No language model grades the review.
Review or settings changed. Score again to refresh this scorecard and enable export.
All scores
| Dimension | Score | What it counts |
|---|
A citation is not an exact span of the contract, quotes fewer than three words, or the review cites nothing. Any blocker rejects the review, whatever its other scores.
A requirement of the release policy is unmet: a missed or unexpected clause, a missed or extra risk, a wrong severity, a duplicate flag, or a flagged clause with no citation inside it. A qualified reviewer decides.
The item matches the answer set. A review with no blockers and no review items is pilot eligible, which still requires a lawyer to confirm every flag and citation.
What the scores rest on
This page runs contract_eval-0.7.0rc1-py3-none-any.whl, the same package as the command line, inside your browser. The exported JSON is the result contract-eval score writes for the same file and case. The answer sets were written by hand for three synthetic contracts and record whose side each severity is judged from. A score measures agreement with that answer set; it is not a public benchmark result. Read the metric reference ↗