ENTRUST AI APEX

ENTRUST AI builds governance tools for AI.

APEX came out of a problem we could not get past: a model asked once hands you its first answer, and neither of you knows whether it is the best it could do.

APEX does not answer once. Your request is recorded as a case and sealed with a hash. Three drafts are written separately, each under a different strategy. Two examiners compare them blind, under random labels and in opposite orders. A new version is built from the strongest parts, and it inherits no approval from the drafts. Then a defect judge examines it while a challenger tries to write something stronger, and a final arbitration compares the version against the challenger and against every draft. The decision to release is made by fixed rules in code, applied to the recorded verdicts — not by a model deciding it is satisfied.

If nothing material survives against the text, you get it with a receipt signed with Ed25519, published under its run ID, and anyone you send the text to can check that receipt at entrustai.co/verify without an account and without taking your word for it.

If something does survive against it, you get nothing but the reasons. A refusal is a real outcome here, not a failure. I would rather hand you three findings than a letter you should not have sent.

It is built for the writing where being wrong costs something: the letter to a regulator, the note to a client after a mistake, the board memo recommending a delay.

Two ways in — entrustai.co/app, or the APEX app inside ChatGPT. Three replies free on a new account.

Happy to answer anything about how the examination actually works or about what it refuses and why.