Maintenance release. Fixes the count of masked items in the trial usage statistics (the value was always reported as 0). Detection, model, masking and restore are unchanged from 0.8.8.

New detection model — the same one the Saklam apps use. It runs 24 % faster, masks fewer harmless words (fewer false positives) and is about 40 MB smaller than the previous one. Model and vocabulary table are downloaded once from saklam.com on first start and kept locally afterwards; detection itself still runs entirely in your browser.

Also new: a rule-based layer alongside the model. It picks up reference and case numbers attached to an anchor word ("Our reference", "Staff no."), company names without a legal form, additional place names, and titles preceding a name. The anchor word stays readable.

New detection model — the same one the Saklam apps use. It runs 24 % faster, masks fewer harmless words (fewer false positives) and is about 40 MB smaller than the previous one. Model and vocabulary table are downloaded once from saklam.com on first start and kept locally afterwards; detection itself still runs entirely in your browser.

Also new: a rule-based layer alongside the model. It picks up reference and case numbers attached to an anchor word ("Our reference", "Staff no."), company names without a legal form, additional place names, and titles preceding a name. The anchor word stays readable.

New detection model — the same one the Saklam apps use. It runs 24 % faster, masks fewer harmless words (fewer false positives) and is about 40 MB smaller than the previous one. Model and vocabulary table are downloaded once from saklam.com on first start and kept locally afterwards; detection itself still runs entirely in your browser.

Also new: a rule-based layer alongside the model. It picks up reference and case numbers attached to an anchor word ("Our reference", "Staff no."), company names without a legal form, additional place names, and titles preceding a name. The anchor word stays readable.