Measuring Temporal Behavioral Drift in LLM Agents: The Constraint Consistency Score (CCS) Methodology

LLM-based agents operating in long-horizon sessions are subject to a reliability failure mode distinct from stochastic run-to-run variance: temporal behavioral drift caused by context window rotation. When compaction events discard or summarize early-session context, behavioral constraints specified at session initialization decay toward zero influence in model generation. This drift is invisible to standard reliability benchmarks that evaluate agents in short isolated runs. We describe the Constraint Consistency Score (CCS), a lightweight session-time metric for detecting constraint term decay, and the ghost term decay mechanism for identifying which specific constraints have been lost. We provide a reference implementation, a proposed measurement protocol, and a case study in the TradingAgents multi-agent financial framework.