Evals
Measure whether an LLM feature works with a test set, code-based checks and calibrated judges, so changes are decided by numbers and not by impressions.
Why evals
A prompt change that fixes one example often breaks three others you did not look at. Reading a few outputs and thinking "this looks better" does not scale and is easy to fool.
An eval (evaluation) is a repeatable test: a fixed set of inputs, a way to score each output, and a single number you can compare before and after a change. It is the unit test of AI work.
1 of 8