Detects feature distribution drift between training and production data using PSI scoring.
Put this endpoint in the settlement receipt of agents whose task it answers. You pay only when the recommendation converts on-chain.