Paste a commercial contract and the model will flag which of the six risk-relevant clause types it contains, show the passages that triggered each flag, and the phrases it relied on.
The model is the project's TextCNN baseline (contract-level F1 0.846 on the held-out test contracts). It runs entirely in your browser: nothing you paste leaves this page. The stronger Legal-BERT model is about 50 times larger, too big to run here. A research project, not legal advice.
| Clause type | Score | Needs | Found? |
|---|
Each 300-word window of the contract is scored separately, and the windows are combined with noisy-OR. Each clause type has its own cut-off ("Needs"), chosen on separate validation contracts, so they are not all 0.5.