An AI model reads a section against its source, and reports.
Two checks, named: accuracy and style. Each finding says which segment it is about, what it thinks is wrong, and what it would do instead.
See it in the applicationHide the screenshot

Reviewing your own work is hard for a simple reason: you have read the sentence four times, you know what you meant, and that is exactly why you cannot see what you wrote.
Two kinds of help, and both leave the decision with you: an AI model that reads a section against its source and reports, and a rule you wrote once that stops a segment being confirmed while it is unsatisfied.
Two checks, named: accuracy and style. Each finding says which segment it is about, what it thinks is wrong, and what it would do instead.

A finding takes you to the cell and gets out of the way: what you type there is your translation. Only a fix the AI model wrote has to be accepted.
If the target has changed since the check read it, the AI model's fix is refused and your edit stands: a fix is a whole target, written against the sentence the check read.
Your dismissals, your accepted fixes and everything still undecided are stored with the file. The report is the queue, and only you clear it: by deciding each item, or by running the check again, which starts a fresh one.
Splitting, joining and moving a section break are refused for as long as the mode is on. Findings are keyed to segments, and renumbering them would point every finding at the wrong sentence.
Mark a termbase or one of your own rules as a check, and a segment that fails it stays unconfirmed — the row is marked, and the failures are listed. Dismiss one and that decision is kept, for that segment, in the file.

They come from your own termbases and from regular expressions kept in libraries you attach per project. A rule is tried against a sample before you save it.
Open the whole delivery and it is one continuous grid rather than a stack of tabs, each file's rows under its own name.

Switch on Checking mode and each row shows whether every figure in its source is in its target, compared by value and read the way each language writes numbers: 1,094,950 shares for 1.094.950 actions passes, and ESRS 8 for ESRS 4 does not. F9 takes you to the next finding, and one you dismiss stays dismissed in the document.
Send for Review writes a file your reviewer opens in Trados, memoQ or bAIbel AV, and Import Review lists every change before anything is applied. Each change comes back unconfirmed, with the reviewer's comments, for you to confirm.