Classifier runs return the winning label and optional probabilities.

{
  "label": "invoice",
  "confidence": 0.82,
  "scores": {
    "invoice": 0.82,
    "receipt": 0.12,
    "contract": 0.06
  }
}

Handling low confidence

• Set a minimum threshold in Processor → Settings.
• When confidence drops below the threshold, responses include "status":"requires_review".
• Combine with human review queues (see /guides/human-in-the-loop/overview).

Use the scores to power routing logic in workflows or analytics dashboards.