Human-in-the-Loop Design

The decision engine recommends. The human decides. This division is deliberate and non-negotiable for high-consequence decisions. The human-in-the-loop design presents the recommendation, the rationale, and the alternatives. The human can accept, modify, or override. Every response is captured — accept, modify, or override — with the reason if an override. This data is the training signal that improves the engine. When the human accepts 90% of recommendations and the 10% overridden have worse outcomes, the engine's logic is well-calibrated. When the human overrides 40%, the engine needs tuning. The accept rate is the calibration metric for the logic layer.