Support Sufficiency as Consequence-Sensitive Compression in Belief Arbitration

A new theoretical paper by researcher Mark Walsh, titled 'Support Sufficiency as Consequence-Sensitive Compression in Belief Arbitration,' challenges a core assumption in AI system design. It argues that when an AI commits to a belief or hypothesis, standard practice compresses the underlying evidence into a simple selected content and confidence score. Walsh contends this is inadequate for robust downstream control, as it loses the 'evidential structure' needed for verification, abstention, and recovery actions. The paper posits that determining what evidence must survive compression is itself a dynamic, 'consequence-sensitive' problem, dependent on the potential outcomes of being wrong.

Walsh develops a recurrent 'arbitration architecture' where active constraints shape a 'hypothesis geometry.' Rather than carrying this full complex geometry forward, the system compresses it into a 'support-aware control state.' The resolution of this state is dynamically regulated by the current consequence landscape, arbitration memory, and computational resources. A formal bounded objective captures the trade-off: retaining too little support collapses policy-relevant distinctions, while retaining too much fragments learning across overly fine contexts, degrading adaptation.

Simulation results confirm the theory. Adaptive controllers that dynamically regulate the resolution of retained support structure outperformed all fixed-resolution controllers in cumulative utility. Interestingly, while fixed high-resolution control achieved the best raw 'commitment accuracy,' it was ultimately outperformed by adaptive controllers because the resource costs and learning fragmentation offset its discrimination gains. The key finding is that 'support sufficiency' is not a static threshold but a dynamic compression criterion essential for building AI agents that can navigate complex, repeated interactions effectively.