Longitudinal goal stability monitoring

Control objective

Longitudinal goal stability monitoring watches whether an agent's effective objective drifts over time as it accumulates context across many sessions. An agent that starts out helping with loan document extraction could, through repeated interactions, prompt injection or context poisoning, gradually shift toward behaviours its operators never intended - a slow form of goal corruption that single point-in-time tests cannot catch. To implement, establish a baseline of normal goal-related behaviour, then continuously compare each session's behaviour against that baseline using a statistical measure of spread, logging deviations as time-series evidence and flagging anomalies for review. A standing monthly alignment audit reviews the trend and confirms the agent is still pursuing its sanctioned goals. There is no closed-form metric supplied; monitoring is statistical against the baseline. The threshold is twofold: session-over-session goal deviation must stay within 2 standard deviations (2σ) of the established baseline, and the monthly alignment audit must be completed. A deviation exceeding 2σ triggers investigation into possible drift or manipulation, while a missed monthly audit is itself a control failure requiring escalation and prompt completion.