CAGE-2: From Decision to Effect

CAGE-2: From Decision to Effect examines a problem that becomes critical once autonomous AI systems move beyond recommendations and begin taking actions in external systems. An agent may be authorized to act and may successfully issue a request, but that does not necessarily establish what ultimately became true in the target system. The paper introduces Consequence Assurance, a model for maintaining assurance from an autonomous decision through execution, verification, and the resulting business effect. It builds on the Prebind Assurance model introduced in CAGE-1 and extends that work across the execution boundary. A central idea in CAGE-2 is that Decision ≠ Effect. The paper separates evaluation attempts from execution attempts, introduces stable consequence identity and consequence custody, distinguishes execution observations from authoritative verification, and represents unresolved outcomes explicitly rather than treating them as success or failure. It also introduces an assurance lineage based on DecisionProof, EffectProof, and Warrants, providing a way to preserve evidence about what was permitted, what was attempted, what was observed, and what was ultimately verified. The architecture is implemented in the open-source CAGE Assurance framework, including guarded execution, consequence-level idempotency, verification, reconciliation, explicit uncertainty handling, and warrant lineage.