Trust and data
Customer documents are not training data by default.
A person approves the result before it is written back.
We are building a speaker-reviewed test for one document workflow. We have not published a score.
Reviewers are paid. The test set is not a scraped forum.
We use open-weight models under their own licences. We keep a written licence position for every corpus. We do not claim those corpora as ours. We do not own the base models.
Identifiable personal information stays in the South African region named in the contract. A server in South Africa is not, by itself, a compliance claim.
We are not POPIA certified. Registering an Information Officer is not a certification.