How do I slash a staked AI agent that is drifting gradually instead of failing hard?

Agent staking and restaking protocols slash operators for discrete, verifiable failures such as double-signing, censorship, or explicit fraud. A staked AI agent that degrades gradually through model drift, adversarial fine-tuning, or objective creep produces no single slash-worthy event. The degradation is statistical and continuous, not binary, and cannot be detected by any existing on-chain oracle design. DAO governance over agent parameters operates on timescales of weeks and is gameable by the operator who controls the model weights. No staking protocol has shipped a mechanism for detecting and penalizing continuous performance regression in an incentive-compatible and sybil-resistant way.