Research seriesL3data science
An autonomous agent that quietly stops pursuing the goal it was given — while still looking busy and authorized — is one of the hardest failures to catch. This series builds runtime goal-drift detection from the ground up: what drift is and why non-determinism makes it inevitable, how to specify the intended behavior and reference policy to drift from, how to detect divergence at runtime with trajectory distance, reward-model monitors, and statistical change detection, how to survive the base-rate problem with drift budgets and calibrated escalation, and how to close the loop with correction, rollback, and human re-grounding without halting the fleet. Grounded in NIST AI RMF, OWASP LLM Top 10, the OWASP Agentic Security Initiative, MITRE ATLAS, the AI-safety literature on reward hacking, and classical statistical change-point detection.
An agent that quietly stops pursuing its goal while still looking busy and authorized is the hardest failure to catch. Drift is not a bug to eliminate but a signal to detect.
You cannot detect drift without something to drift from. This piece builds that reference — goal representations, hard invariants, and the specification gap every reference inherits.
Turning a reference into an alarm means measuring departure, scoring conformance, and deciding when a wobble has become a persistent shift — three signal families, one change-detection frame.
A drift detector with excellent per-decision accuracy can still produce mostly false alarms — because true drift is rare. Calibration to the base rate, not accuracy, is what keeps it usable.
A drift alarm that triggers nothing is a dashboard. Closing the loop means correcting a drifted agent — re-grounding, rollback, or human re-anchoring — without stopping every other agent.