Ristmark is built so you never have to guess whether to trust the output. Two halves make that work.
Structure, layout, formatting and rendering are handled entirely by code — never AI. The same input always produces the same result. This is why generation is safe to automate: there’s no model in the generation path to drift.
At migration time only, the AI does one job: decide which spans of text are variable and label them. It never touches structure or layout. For every field it reports a confidence score, banded as:
The point isn’t that the AI is never wrong — it’s that when it’s unsure, it tells you, so your review time goes to the handful of fields that need it instead of re-reading the whole document.
When you correct a field, Ristmark records it. The next time it sees the same thing, it applies your correction automatically — and shows you that it did, so nothing changes behind your back.
Reliable field detection depends on a document’s real structure. Word and RTF carry that structure, so they’re the primary, highest-fidelity path — confidence scoring is strongest there. PDF and image sources are also supported, via a visual reconstruction (Beta) mode: Ristmark reconstructs the layout visually instead of reading real structure, which is inherently lower-fidelity, so review those documents more carefully before you rely on them.