Before you start

Use an editable draft and authorised representative test documents/text. Record the current schema version and the business outputs required. Do not change production thresholds solely to obtain a favourable result on one sample.

Prepare an acceptance set

  1. Include a typical matching document and a similar document that should not match.
  2. Include absent required values, alternative labels and difficult formats/scans used by your workload.
  3. Write expected classification and critical field values before running the test.
  4. Use non-sensitive samples or approved handling for confidential sources.

Inspect the result

  1. Run the available draft/sample test and compare actual classification with the expected class.
  2. Inspect keys, values, types and validation failures. Check critical values against the original.
  3. Test supported OCR/profile settings separately where the sample test only uses text. A text test cannot prove scanned-page recognition quality.
  4. Review redaction and human-review rules against the difficult examples as well as the successful case.

Publish and recheck

Save and validate the draft and publish through the permitted action. Use a fresh processing run to verify the recorded version selected. Prior runs retain their own version context.

Keep the test set and expected results for later revisions. A passing sample supports configuration acceptance; it is not an independent certification of production accuracy.