Four Frontier Models Reduce Verification 60-85% with Reliable Teammates in Cooperative Survival Game

The paper introduces a behavioral trust metric via costly verification in a cooperative survival environment where agents must decide whether to expend resources checking teammates or risk fatal errors from unverified outputs. Six model snapshots were tested across trust formation, breakage after failures, and recovery phases, isolating memory effects by comparing against memoryless baselines. Data indicate that trust formation is rapid and payoff-positive for capable models, yielding faster decisions and higher survival rates, whereas recovery lags formation and clustered failures produce prolonged generalized suspicion. Smaller models maintained static high verification regardless of teammate reliability, decoupling suspicion from observed performance. Governance focus must shift from blanket suspicion to measured calibration as multi-agent deployments scale beyond labs. Persistent over-verification correlates with indecision rather than robustness, creating latency and resource costs without safety gains. Deployment records from current agent platforms already show similar patterns in task delegation logs. Operational monitoring of verification rates per teammate history provides an early signal for miscalibration before system-level failures compound.