What you can check.

Every flag shows its page and a label that says how sure. This page says what the labels mean, how we test, and what is still being tested.

Open a flag in the demo

Five words, plus one for when it fails.

Nothing is marked verified that is not. A label never changes while you watch.

EXACT: The quoted sentence was found on this page, word for word.
NEAR: The sentence matches the page closely, with small differences.
LOCATED: The page is right; the exact words were not matched.
VISUAL: Read from a drawing or picture. It carries a region, not a text match.
INFERRED: Drawn from several places. It points to the pages it used.
UNVERIFIED: The text could not be matched to the page. Check it by hand.

What we measured.

We do not publish accuracy figures yet. When a test passes its target with enough checked items behind it, the result appears here with its date and its limits. Until then, the labels above and the list below are what we stand behind.

Public packages, real quotes.

We test on 23 public bid packages (36 documents, 8,612 pages) from agency procurement sites and public records. Answer keys are real quotes plus edits we plant, and results are scored independently of NIX’s own labels.

Still being tested

We make no claim on these until they pass.

Page-by-page coverage reporting on very large packages.
Closing every item on a trade checklist, so nothing is left open.
Scanned and rotated drawings.
Citation exactness on some packages.

What NIX does not do

It does not take off or price.
It does not tell you whether to bid, and it gives no win probability.
It does not replace your estimator. It is a second reader.

Where your documents go

The security page says it plainly. For a security questionnaire, email nshulman@natex.ai.