Metrics must match the business task
Accuracy alone may not be enough. The right metric depends on the cost of false positives, false negatives, latency, throughput and manual review.
Before implementation, the client and delivery team should agree how results will be measured and reproduced.
Common acceptance elements
- Representative test samples and data source description.
- Target metrics, tolerance ranges and edge cases.
- Runtime environment, performance requirements and fallback method.
Why boundary cases matter
Real production data often contains rare classes, noise, missing fields or environmental changes.
Recording known limitations helps operations teams decide when to trust the algorithm and when to trigger manual review.