Revelation Control Optimizes Information Disclosure in Learning Systems

This research introduces "Revelation Control," a theoretical framework designed for learning systems to manage the disclosure of hidden states. The core problem it addresses is how to choose priced interventions that reveal information only to the extent that it can influence a significant decision, while simultaneously quantifying the direct utility or progress generated by the intervention itself. The framework defines concepts such as decision-sufficient revelation and revelation depth, integrating static Bayesian refinement with state-dependent continuation values. It provides an exact cost-adjusted factorization criterion: a new piece of information is considered decision-nonredundant only if states that share a scalar summary fall on opposite sides of a predefined Stop/Continue decision boundary, considering the intervention's cost. Experiments conducted with Qwen2.5-7B and Mistral-7B-v0.3 demonstrated that deeper future-learning probes offer positive decision value, and productive reuse of information yields clear utility advantages for equal computational effort. Qwen showed evidence for a shallow revealability regime, while Mistral's scalar continuation architecture maintained positive lower bounds, suggesting structural rather than purely numerical transferability of the decision theory and evaluation protocol.