Redaction is a measured detection pipeline with residual risk

Specify which entities and combinations matter for the project: contact details, account identifiers, authentication secrets, health-related narratives, or other context-dependent information. The same string may be a fictional example, a public organizational contact, or a person's private record. A regular expression can propose number-like candidates; checksums and surrounding words can improve precision, while named-entity recognition can find less structured spans. None of those mechanisms understands every context or language. Include misspellings, split identifiers, unusual formats, quoted text, and image-derived content in evaluation when the corpus actually contains them.

The current Presidio primary repository describes customizable recognizers combining patterns, contextual rules, checksums, and learned recognition, while explicitly disclaiming complete sensitive-information detection. That limitation should shape the workflow rather than appear only as fine print. Define whether an error is counted per span, entity, document, or person. A detector with strong span recall can still miss one critical identifier in many documents. If an annotation policy changes what is sensitive, previous recall estimates do not automatically measure the new task, and unlabeled sensitive content must not be counted as true negatives by default.

Use independently reviewed examples covering the intended sources and categories to estimate misses and false alarms. Inspect rejected or supposedly clean material as well as flagged spans. When a detection engine is reused as its own judge, both can share a blind spot and create an artificially reassuring score. Validate exact span boundaries and output rendering after replacement, including overlapping detections and Unicode offsets. Removing a name while keeping a rare occupation, location, and event can leave an identifying combination, so detected-entity replacement is not automatically a formal anonymity guarantee.

Under an explicitly assumed miss probability, expected missed count follows linearity of expectation; independence is not needed for that expectation when the marginal probabilities are correct. The probability of at least one miss does need a joint model, and an independent approximation can be badly wrong for repeated templates. Two detectors do not provide multiplicative improvement merely because they run separately: shared training, features, or format limitations correlate their failures. Report the scope and uncertainty of residual estimates, and avoid saying a corpus is fully anonymized just because recognized strings have been replaced.

A release boundary includes derived text, source URLs, document metadata, thumbnails or images if included, debugging samples, search indexes, and cached intermediate files. Replacing the visible article's phone number is insufficient if the same number remains in a filename or error log. Track versions and access boundaries for raw and sanitized representations, and test that consumers receive the intended one. A deterministic hash can support matching but may permit guessing or membership tests, so hashing an identifier is not automatically equivalent to removing its informational value or resolving the source's use conditions.

The probability lab isolates an assumed independent miss process. Reset, explicitly set BOTH miss probability p=0.1 and n=10, predict P(K≥1)≈0.6513, then Play or Step. Reset and re-set BOTH p=0.1 and n=20; predict about 0.8784. Edits retain t and playback highlights exact counts without scanning anything. These controls do not implement redaction, contextual judgment, re-identification attacks, correlated detector failures, or privacy certification. The increasing chance of some miss illustrates why a seemingly strong per-item detector still needs careful release-level assessment.