01Freeze the sample
Define the workflow, entities, material claim types, countries and measurement window before examining results.
02Record the baseline
Measure the unverified workflow first, including unresolved entities, unsupported claims, corrections and downstream decisions.
03Review disagreements
Route a reproducible sample of model and source disagreements through documented human review instead of treating the API as ground truth.
04Publish denominators
Report the sample size, exclusions, time window, definitions and customer approval status beside any outcome percentage.