Every claim about an AI system should be checkable by someone who doesn’t trust the people who made it. We build that standard into everything we ship — starting with a receipt for every action, verifiable offline, against a verifier we don’t control.
This receipt is the lab’s standard, live. Don’t take our word — break it. One byte and the math fails.
If we can’t prove it, we don’t claim it. If we tested it, we publish it — including the failures.
Every action an agent takes is signed (ECDSA P-256) and hash-chained (SHA-256) into a receipt you check offline, with an open-source verifier. Flip one byte and verification fails. Neutrality is the property: the thing that checks the system is not part of the system.
The open experiment: a learner that adds capability without erasing what it had. Today it survives toy-scale curricula — arithmetic, not the real world. That test is still ahead of us, and we will publish the result either way.
Where both programs meet the real world: an AI that does whole jobs, signs every action, and hands the proof to you. Free to start; the verifier is free for everyone, forever.