Limits and evaluation¶
Use version 0.2.0 for supported identifier candidates and email syntax. Review the source because unmatched personal information can remain.
Coverage¶
The country-specific types are AR_CUIT (CUIT/CUIL), AR_DNI, CL_RUT (RUT/RUN), CO_CEDULA, CO_NIT, DO_CEDULA, MX_CURP and MX_RFC, plus EMAIL. Protective AMBIGUOUS_ID emissions remain unverified; they do not establish nationality or another validated identifier scheme.
Names and addresses are unavailable. Phones, bank/card details, unsupported identifier schemes, OCR corruption and unusual spacing/masking are not comprehensively detected. The image locks its identifiers-only profile; non-off neural overrides are rejected.
Validation boundaries¶
Implemented checksum arithmetic is narrower than detection. CURP and Dominican cedula checksums are unverified; RFC arithmetic is partial. A syntactically plausible or checksum-valid value does not establish assignment or identity. Email syntax does not establish ownership or deliverability. No issuer lookup, identity decision or compliance certification is performed.
Measured scope¶
The retained V4 evaluation contains 600 post-freeze AI-authored, frame-based fictional documents. It produced 600 exact document results under its declared targets, with no expected-value leaks across mask/replace/hash. These are correlated synthetic diagnostics with no human review; they do not establish customer or population accuracy. Earlier failed and exposed evaluations remain disclosed.
Read the model card for exact-image/AWS measurements, country-by-country boundaries and prior failures. Read the sources separately from the data and software-license review. Evaluate your own authorized workload before relying on the redaction.
Last updated November 15, 2026 · Suggest an improvement