How detection works

How Lex Cloak finds sensitive information

Detection runs on your computer, with a rule-based engine plus a local model. Lex Cloak reads every page the way a person sees it, using OCR even on PDFs that already have a text layer, because that layer can be missing, misaligned, or hidden. The engine looks for known patterns of sensitive information, scored by confidence, and surfaces them as red and amber boxes for you to review.

Two things to know up front: the goal is to help you review faster, not to make the decision for you, and the engine errs toward over-detection so the cost of a miss falls on review time rather than on a leak.

Category chips

Category chips in the sidebar group matches by type, such as Name, Date of birth, Social Security number, Address, and so on. Each chip has a count and a toggle. Turning a chip off removes those matches from review AND from the saved file — the saved export reflects the categories that are on at the time you download, which is why the pre-download checklist surfaces categories you turned off as a finding worth a second look.

Use the All and None buttons at the top of the section to turn every category on or off in one click. Keyboard: Space toggles a chip and the arrow keys expand or collapse it.

Confidence and uncertain matches

A solid red box is a confident match and a dashed amber box is an uncertain one that is asking for your decision. The Confidence section of Settings has a slider and a Show uncertain matches checkbox. Together they change how many uncertain matches you see for review, not what counts as sensitive. Think of these as review-visibility controls, not redaction-strength controls.

A lower threshold surfaces more borderline items, which is useful when a document type rewards extra caution; a higher threshold keeps the review list shorter on documents whose patterns are familiar to the engine.

Detection profiles

A detection profile tunes the engine for a kind of document. The dedicated detection profiles page covers what each profile changes and when to switch. The toolbar template chip shows the active profile and how well it fits the document; switching profiles re-runs detection.

What detection does and does not cover

Detection covers many common patterns of sensitive text, but it does not catch everything. Unusual formats, complex layouts, identifiers in languages the engine has not learned, and anything inside an image without legible text can slip past. Signatures and handwriting are not detected automatically on purpose — see signatures and handwriting.

Repeating headers and footers are detected and redacted as bands by default, while page numbers are kept. The dedicated headers, footers and page numbers page covers the toggle and the symptom of my letterhead disappeared.