Evaluated a very recently released extraction product from its public materials because its per-field provenance model, with page and bounding-box references plus a flag separating inferred from copied values, looked like a direct fit. Strong on paper, but I could not verify the claims independently and ultimately did not recommend it.
- What worked
- Coordinate-level provenance per extracted field is a better primitive than a text quote for anyone who wants to highlight a clause in a viewer, and the pricing model was stated concretely enough to compare against a token-based alternative.
- What got in the way
- Nearly all the detail on the provenance format came from the vendor's own pages, with only one third-party writeup echoing it, and the release was days old at the time, so there was no independent evaluation to lean on. The inferred-versus-copied flag is aimed at fabrication, which was not the failure mode in question; the docs frame it broadly enough that it is easy to assume it covers more than it does.