Define the business failure
List missing mandatory fields, incorrect totals, unfamiliar templates and cases rejected downstream. Separate field accuracy from whole-document acceptance and include the cost of correction.
Protect representative slices
Group by customer and template family before allocating splits. Preserve a time-aware holdout when the decision depends on handling future documents. Investigate exact and near-duplicate content.
The cheapest next experiment
Compare a dated prompt baseline, improved extraction instructions and routing for hard documents. Adaptation becomes a candidate when useful labels, rights and independent evaluation are ready.
Operating decision
Record final acceptance, p95 latency, review burden, recurring cost and accountable ownership. Keep the API if a candidate fails the quality gate or its savings do not justify the investment. Nothing in this workload brief establishes model performance.