Constat AI

the problem

A week, answered from memory.

A buyer sends a security questionnaire and it lands on whoever is least able to refuse it. They answer from memory and from last year's spreadsheet, the deal waits, and nobody can say afterwards which claim came from which document or whether it was still true on the day it was sent.

what I built

Three things, and the third is the one that matters.

It publishes a trust page, so the questions a buyer can answer for themselves never reach a person at all. What is left it drafts, and every draft carries a citation to a versioned, approved source rather than to the model's memory.

And where nothing in the evidence supports an answer, it refuses, names the gap, and routes it to whoever owns it. Anything can print a confident paragraph. Declining to is the harder half, and it is the half that makes the other two worth having.

how it stays honest

The gates are code, not judgement.

The refusals are not the model being cautious. They are deterministic gates that run after it, on the structure of the evidence: a certification roadmap is not a certificate, so an answer claiming the certificate never ships, and the capture on this page catches exactly that refusal being made.

The demo you can run says on screen that it works from a sample evidence pack and calls no API, and the caption under every capture here repeats it, because a product whose argument is that an answer must carry its provenance cannot be shown without carrying its own.