We do not ask you to trust us. We show you the measurements.

The scorecard, as CI produces it

The numbers per entity

From our gold sets, in hermetic replay. Everything here is written by the quality gate itself, not by hand.

What the numbers mean, in plain words

• Recall: Of 100 personal data items actually present in the text, how many we detected. 1.000 means none slipped past us in the set.
• Precision: Of 100 items we masked, how many really were personal data rather than a false alarm. 1.000 means no needless masking.
• F1: The two combined into one number, so that a system that simply hides everything does not look good. 1.000 is perfect.
• Leak recall: How many of the items that had to be hidden were covered at least in part. It is the number we care about most: one uncovered item is a leak.

Every figure refers to our own evaluation sets, not to every possible document. Always review the findings before you send a text.

The other side of the same problem

Not masking what does not need masking

We would rather mask one item too many than miss one (the “recall-first” principle). That carries an obvious cost: a system that masks everything never misses anything, and is useless. In September 2026 we measured and reduced false findings, under the explicit condition that recall was not allowed to drop by a single value.

• False findings on an adversarial set: 241 → 92 (62% reduction)
• Documents with no false finding at all: 64.1% → 86.9% (+22.8 points)

• Recall on the Greek reference set: 1.000, unchanged at every step
• Overall F1 (micro): 0.994
• The byte-exact reference snapshot stayed identical throughout
• Two candidate rules were rejected because the quality gate priced them

Language coverage

Three tiers. For each one we say exactly what holds.

Production depth

Greek · English

Full detection with Greek-specialised models and deterministic validators: AFM checksum, AMKA Luhn, IBAN, land-registry codes, IDs, plates. Name inflection handling and identity linking across documents. Validated on real public documents, not only curated text.

Measured coverage of names and structured data

German · French · Spanish · Italian · Portuguese · Dutch · Polish · Romanian

Names via a multilingual model at 100% recall on our evaluation sets, including inflected Polish forms. Name lists and surname morphology, plus deterministic dates, postal codes, phone numbers, VAT and IBAN.

Other scripts: flagged before sending

Cyrillic · Arabic · Chinese and other scripts

Names written in Cyrillic, Arabic or Chinese script are not detected yet. When a text contains them, filterit flags it before anything leaves, so the decision stays with you. Script-independent items such as IBANs, emails, phone numbers and card numbers are detected as usual.

How we measure

The methodology is the product. These are its four instruments.

01 Quality gates that only tighten

Labelled evaluation sets per language and data type. The recall floor is 100%: a single leak in any set automatically blocks the release. Thresholds ratchet upward and are never silently relaxed.

02 Real-document validation

Curated sets are not enough. Reality has stamps, scanned forms and captions printed under the values. We test on 22 real Greek public documents with labelled data, and every finding becomes a permanent rule with a permanent test.

03 Deterministic first

Anything with structure is verified by proof, not statistics: the Greek tax number by its checksum, AMKA by Luhn, IBAN by mod-97. Neural models handle only what cannot be proven, such as names and addresses, and are wrapped in confirmation layers.

04 When we cannot, we say so

Unreadable handwriting? An unsupported script? The product does not pretend. It shows an explicit warning before anything leaves. Silent failure is the only truly dangerous failure mode in a privacy tool.

What gets detected

34 categories, 11 of them specific to Greek documents

Greek identifiers are validated (check digit, format, context) rather than matched as digit runs, and names are followed across grammatical cases. Tap a category for details.

In scanned PDFs and images, signatures and stamps are covered by visual redaction, with a preview before it is applied.

Every category and the accuracy measurements

How to read the measurements

The figures on this page come from our evaluation sets and from the real documents we test on, and describe how the system behaves on those. On a new document performance can differ. That is why the privacy panel shows what was masked in every conversation, visual redaction goes through a preview before it is applied, and the system tells you wherever it cannot read the text. The final decision is always yours.

Measurements last updated: September 2026

See it on your own document

Upload an invoice or a letter and see what gets masked, before anything leaves.

Try it free