Evaluate an AI change with real tasks and visible failure criteria
Keep a representative test set, define what counts as success and compare a new prompt or model without changing the test at the same time.
In brief
An evaluation asks whether an AI change improves a defined task. It needs examples and a scoring rule; a persuasive answer to one easy prompt is not enough.
Start with the decisions and errors that matter to your actual use, not an unrelated public benchmark.
Choose measurable criteria
For receipt extraction, check the amount and currency against the source. For a support guide, check that the steps match the platform and actually address the symptom. For tool use, check permission and duplicate-action handling. One overall “good” score can hide a serious failure.
Keep comparison conditions stable
Use the same examples, output requirements and available evidence for each candidate. Record the prompt and model versions. Separate tuning examples from a held-back check so memorizing the small test does not look like general improvement.
Review failures, not only averages
Inspect which cases got worse and why. Use deterministic validation for fields or rules where possible, and human source checks for meaning. Report the tested scope and uncertainty rather than claiming the model is universally better from a narrow dataset.
Before you act
Model judges can have biases and repeat model mistakes. Important factual or safety cases need evidence-based human review and appropriate automated checks.
What to do next
- Collect representative permitted examples, including difficult and absent-answer cases.
- Define success and specific failure categories before comparing versions.
- Run comparable configurations and inspect regressions, preserving a separate set for later checks.
