AI agent monitor scoring flawed: coin flip beats standard metric

A developer found that the standard F1 metric for AI agent monitors rewards early detection so heavily that random guessing achieves a 0.88 F1 score. After fixing the metric to only count detections on actual drift steps, the coin flip dropped to 0.19 F1. The corrected baseline shows GPT-4o-mini judge achieving 0.672 F1, while production verifiers trade off recall for low false positives.