THE agents

A team of AI agents that has to prove every change it makes to itself. After each run it proposes one change to its own rules or setup; a plain-code gate keeps it only if it helps on practice tasks and doesn't hurt separate checking tasks, then every kept change is removed again on fresh samples and deleted if the score doesn't fall. On a GLM-5.3-Flash builder, it raised the score on unseen tasks from 19% to 94% (one run, 9 unseen tasks), mostly by fixing replies that ran out of room. On Nemotron Nano it kept nothing it couldn't prove. Built on Crusoe Managed Inference with three models (builder, verifier, proposer). All code was written at the event.