Redact’s real question about each line of a page is small and fixed: does this contain a name, an email, an account number? A decision model is built for exactly that kind of question. Laya doesn’t write anything. You give it some text and questions you wrote in advance, and it returns a probability for each answer.

​
Why typed questions fit

The answer has a fixed shape: one probability per question per line. There is no free text to parse, no output that can be malformed, and nothing for the app to repair. The model can’t invent a line that isn’t on the page either, because it only ever answers about the lines it was given. The probability comes from the model’s own scoring of each option, not from a number it writes. That makes a threshold meaningful, and it would make a sensitivity slider honest. Laya also gives the same answers on every run, with no sampling settings to pin. It is quick, too. On Redact’s labelled pages, Laya decides a 12-line page in under a second, while OCR on the same page takes 20 to 30. A step that cheap can run twice, which is what makes the copy check in lesson 11 affordable.

​
How the threshold works

Redact asks one noul question per field, such as “This text contains a person’s name.”, plus one built from the typed direction. For each line it takes the most likely question and keeps it only if the probability is at least 0.5. Lower thresholds catch more fields and ask the reviewer to reject more; higher ones show fewer, surer suggestions. Decision questions lists every question and the exact rule.

​
How you ask matters

A single multiple-choice question failed badly in testing. Asked “which kind of information is this?”, with none among the options, Laya answered none at 98% for a name, an email and a street address. Asked one yes/no statement per field, the same model caught all six fields on the test chat. Decision models reward statements a person could check against the text, one fact at a time.

​
Where it stops

On the four evaluation pages Laya finds every labelled field on three and adds one suggestion the labels didn’t expect: “keep the amount confidential.”, flagged by the direction question. Its F1 scores are 0.83 to 0.92. For a privacy tool that’s the safer kind of error: an extra suggestion costs one click on Reject, while a missed field is something the reviewer has to notice alone. Laya can only answer the questions it’s given. It can’t explain a suggestion in its own words, summarise a document, or spot a kind of sensitive field nobody thought to ask about. Those would need a generative model, and the human review in About Redact’s trust boundary still matters either way.