Check claims before you rely on them

Mathematics with a proof assistant, code with tests, facts against sources you trust, and other claims with a second model. Anything that cannot be checked goes to a person.

Each step gets a calibrated score of how likely it is to be right. You set the threshold below which a step must be checked or reviewed.

Checking answers that come from another system is one of the workflows we can build an agent for, as long as the answer can be broken into claims.