Ovrin publishes no accuracy figure that its own evaluation harness cannot reproduce — and today it
publishes none at all.
v1.0.0 promises the API, not accuracy
Ovrin is at v1.0.0, and that version number is a promise about compatibility: the API will not break
without a v2. It is deliberately silent about how accurate the library is, because that has not been
measured. The two claims are kept apart on purpose.
Where evaluation stands
- The corpus is synthetic.
eval/corpus/holds 25 documents, five in each of five categories —
invoices, receipts, forms, statements and identity documents — and every one was generated. - No run has been committed. The only report is
no-run: zeros from the report renderer over an
empty set of results, not a measurement. - Confidence is not calibrated. It is a ranking signal until there is a real corpus to calibrate it
against.
Why synthetic documents are not enough
A generated corpus measures the generator as much as the extractor. It lets the harness run, and lets
the harness itself be tested, but a figure computed on it would say little about the documents you care
about.
What would change this
A corpus of real, redistributable documents in every category, committed reports across at least two
provider generations, and a calibration run. Those are the open items on the
roadmap.