Your pipeline sends an extraction to human review when the model's self-reported confidence for any field is below 0.8. An audit finds that fields reported at 0.9 or higher on handwritten intake forms are wrong 25% of the time, while typed forms at the same confidence are almost always right. Reviewer capacity is fixed. What should you change?

Answer: C — Calibrate field-level confidence against a labeled validation set for each document type, and set routing thresholds from those calibrated results.

Raw self-reported confidence isn't reliable, and the audit shows it means different things for different document types. Section 6 calls for field-level confidence calibrated with labeled validation sets, with review thresholds set from that calibration, so that limited reviewer time goes where errors actually happen. (a) is a prompt tweak to a number that isn't calibrated in the first place. (b) averages away the weak field signals that matter. (d) ignores the fixed capacity constraint.