Continuous behavioural monitoring beyond static evaluations

Control objective

Continuous behavioural monitoring recognises that passing a pre-deployment evaluation is not proof an agent will behave the same way in live production, where real users, edge cases and adversaries differ from the test set - a gap that lets risky behaviour emerge only after launch. This control compares what the agent actually does in production against how it behaved in evaluation, catching drift, degradation or manipulation that static tests miss. To implement, capture the agent's evaluation-time behavioural profile as a baseline, instrument production to measure the same behaviours on live traffic, compute the deviation between live and eval behaviour, and feed anomalies into review; complement this with a recurring red-team exercise that actively probes for new weaknesses, logging results as evidence. No formula is given; monitoring is comparison against the eval baseline. The threshold is that live-versus-eval behavioural deviation stays under 15% and that a monthly red-team is completed. A deviation of 15% or more triggers investigation into why production diverges from tested behaviour, and a missed monthly red-team is itself a control failure that must be remediated, since it leaves emerging risks undiscovered.