AI assistants increasingly act on external systems, and a confident but wrong call can be costly. Most safeguards allocate scrutiny by generic uncertainty, treating a reversible read and an irreversible transfer the same.
CASEL sits between the AI and its tools. For each action it estimates failure probability, consequence, reversibility, ambiguity, and evidence freshness, then picks among direct execution, verification, sandboxing, clarification, approval, or abstention by the cost of each mode. In controlled simulations, all methods had the same options and compute cost.
Harmful side-effect trajectories fell from 14.75 percent under uncertainty ranking to 7.02 percent. A simpler risk-ranking matched the harm but sent many more actions to human approval. Reversibility-aware routing mattered. Effectiveness was simulated, not measured in deployment.