Skip to main content

File-wide checks

Most checks look at one segment. These read the whole file and report a problem that spans several segments as one finding, on the first segment to fix. They are rule checks: they use no AI words and run with AI checks off. A package of files is read as one file.

In the report​

The finding shows Also in with the other segments to fix and Used elsewhere with the segments that use the form the file uses most. Click a segment number to open it. The finding is Found by rule check. It counts once in the score, however many segments it names.

What they find​

  • Term translated two ways: a source term seen three times or more, translated one way in most segments and another way in at least one. Hint: backup disk as yedekleme diski and yedek disk. The same source sentence translated two ways stays with Inconsistent translation.
  • UI label translated two ways: a label in quotes, in a tag or in a path with > written two ways. Hint: Settings as Ayarlar and Seçenekler, or Settings > General as two paths.
  • Heading style mixed: headings in Title Case in some places and in sentence case in others. Not run on scripts without capital letters.
  • Step numbering: the source numbers its steps 1 to 5 and the target repeats, skips or reorders a number.
  • Punctuation or format mixed: list items that end differently, two quote styles, full and half width punctuation in Chinese and Japanese, two date formats (23.10.2026 and 23 Ekim 2026) or two ways to group digits (1.000 and 1000). Only where the source is consistent.
  • Name or unit written two ways: a product name or a unit written two ways (Northwind Pay and NorthwindPay, 5 GB and 5 gigabayt). A segment the brand checks already flag is left to them.
  • Sentence in two segments: the same target sentence in a segment whose source does not have it.
  • Form of address mixed: siz and sen in Turkish, vous and tu in French, usted and tú in Spanish, 您 and 你 in Chinese. Quotes and dialogue are left out. German has its own check.
  • List item in another form: in a list of four items or more, one item written in another form, such as an instruction among noun phrases.

Switching them​

Checks, File-wide checks turns them all on or off for a profile. They are on by default. Each one also has its own row under Severities, at Minor by default; raise one there.

Whole-file review​

On the Solo plan and up, the AI can also read each file as a whole, after the segment review. It looks for the same kinds of problem across segments: a term, a UI label or a name written two ways, headings in two styles, a list item in another form, a person or role name that changes, and repeated content. It never reports an error one segment shows on its own; the segment review does that.

  • In the report: each problem is one finding, Found by whole-file review, on the first segment to fix, with Also in and Used elsewhere like the checks above. The report opens with one line, Whole file: 3 findings, and its i says what was read. Origin, Whole file in the filters shows only these findings.
  • No repeats: when a rule check or the segment review already found the problem on that segment, that finding gets the other segments under Whole file: also in instead of a second row.
  • Score: a finding counts once, however many segments it names, and at most as a major.
  • AI words: none extra. The file's words are counted once, for the segment review.
  • Long files are read in parts that overlap a little, with one pass at the end that joins what the parts found.
  • Packages: each file of a package is read on its own.
  • Your text as data: the text of your document is sent as data between delimiters, and an answer that repeats our instructions is dropped, as in the segment review.

It runs only once its checks are measured: each kind of problem is switched on for the paid plans when it is right at least 7 times in 10 on our document test set. Until then the switch below changes nothing.

Settings, AI review, Whole-file review turns it on or off for the workspace; a profile can say on or off under AI review. With AI checks off it does not run.