FAQ

Does any document leave my machine? No. There is no code path that sends document text anywhere. The only network access in the project is dueproof fetch-model, which you run by hand once to download translation weights.

Does it use ChatGPT or another LLM? No. Extraction is rules you can read and change; translation is a small, dedicated model (IndicTrans2, about 200 million parameters) running locally.

How accurate is it? Measured on the project's small fictional corpus: every obligation it reports is real, it finds about three quarters of them, and every due date it reports is right. On your documents it will differ — measure it with dueproof eval. It is built to be wrong visibly: an unlocatable passage gets no box, and a doubtful translation is flagged.

Can it read handwriting? No.

Does it need a GPU? No. Translation runs on a CPU; a page takes a second or two to read and a few seconds to translate.

What does it cost? Nothing; it is Apache-2.0 licensed. Hosting, support and integrations for organisations are planned.

Why "Dueproof"? Every due date, with proof.