When Can Safe Controllers Adapt? Information before Commitment

What the math gives to ML

The paper gives a causal way to identify when safe adaptation is impossible: the controller must decide before an action that would eliminate a safe continuation, so information arriving from that action cannot justify the decision. Its transferable asset is the stopped, precommitment KL divergence, which separates evidence available at the decision time from evidence obtained only after taking a potentially unsafe experiment. This can become a safety-aware exploration gate for neural world models or policies: commit to aggressive model-specific behavior only when the current history contains enough distinguishability between plausible dynamics, and otherwise remain conservative while explicitly tracking the unavoidable oracle gap. The framework is especially useful for diagnosing whether poor adaptation is an algorithmic failure or an information-theoretic limitation.