Omit Redact 1.0.0 beta 7

Seventh beta of Omit Redact. Fully offline PII detection and redaction for Windows.

A correction, first

The Accuracy tier has been returning nothing on real documents since it shipped. It had no chunking, and transformer attention costs grow with the square of the input, so anything past roughly 10 KB of text ran out of memory and came back with zero detections. A 28 KB text file died after 52 seconds trying to allocate 17.5 GB. The run reported success. Because Accuracy ships only in the offline installer, and because the failure was silent, every offline user redacting a document of ordinary length got nothing back and no warning that anything had gone wrong.

It is fixed. Text is now processed in overlapping windows, with the overlap de-duplicated so nothing is counted twice and nothing falls in a seam. Anything at or under one window is handled exactly as before. That same 28 KB file now returns 407 detections in 51 seconds.

A PDF that never finished now takes about three seconds

Any PDF containing a table border or a rule under a heading was routed through image recognition, on every page. A one point rule is two pixels tall, and the recognizer scales the short side up before reading, which turned a thin line into a 286 megapixel image. A six page text-layer PDF took over thirteen minutes without finishing and reached 20 GB of memory, at which point the machine is swapping rather than working, so it looks idle while getting nowhere.

Image recognition now runs only when a page actually contains an image. On our test document:

	Before	After
Six page PDF	over 13 min, never finished	3.2 seconds
Peak memory	20,004 MB	57 MB

It also became more accurate, not less. The eight findings that disappeared were all misreadings of text that was already readable directly, such as 1.brandtner@ for l.brandtner@. No real personal data was lost.

Redact needed two clicks, and the review sometimes never appeared

Pressing Redact could do nothing at all. Switching between Files, Text and Email, or moving to Settings, discarded whatever was in flight: typed text, a scan waiting for review, the run itself. The result was a dead-looking button and a review you never got.

A run now survives leaving the screen it started on. Start a file redaction, go to Settings, come back, and it is still going or waiting for you. There is also a progress bar along the bottom of the window that keeps reporting from wherever you are, including which page of a PDF is being read.

The block list applied to the clipboard and nothing else

Terms entered in the block list are described as values that must never leave a document. That was true of the clipboard guard and untrue of every file, because the file, text and email screens were sending a freshly built default policy instead of the one you saved. Per type choices such as mask, hash or keep were discarded the same way. Both now reach every run.

Protect forgot its settings every time you opened the window

Protect keeps working after you close the main window, which is the point of it. But opening the window again overwrote its settings with defaults about half a second later, silently, without you touching anything. Its configuration is now read back from disk on open, and nothing is written until you actually change something.

New: Verify

Give Omit a file and the values that were supposed to have been removed from it, and it tells you whether they can still be recovered. It works on files any tool produced, not only ours, and it is free with or without a licence.

It reports three outcomes rather than two, because a check that could not run is not a pass. If part of a file cannot be examined, an image with no text layer for instance, it says so plainly instead of calling the file clean.

Also in this build

• The Accuracy tier was a paid capability that was only enforced on the clipboard lane. Files, text and email accepted it without a licence. It is now enforced everywhere, and a run that has to fall back to the faster model says so instead of quietly using it.
• A vault carried over from an older install could come back empty. The upgrade copied one file of what is a multi-file database, so anything not yet folded into the main file was left behind, and the check afterwards compared that file's size to itself. Every token would have become permanently unrecoverable. The copy is now made through the database engine and verified by counting rows.
• The detection model install was trusted on the basis of a marker file. A broken install therefore reported itself as present and then failed on every call, returning zero findings and reporting success. It is now verified properly, and a failure falls back to the standard model and tells you it did.

Pick an installer

	Download	On disk	What you get
Offline	931 MB	1.5 GB	Every language, plus the Accuracy model. Never contacts the network.
Light	167 MB	517 MB	English and the Fast model. Other languages download on demand. Accuracy is not available.

Before you install

These installers are not code signed yet, so Windows SmartScreen will warn you and you will need to choose "More info" and then "Run anyway". That is expected and it will stay that way until a certificate is in place.