Evaluation (Evals)
Evaluation is measuring an AI system’s output against known-good cases, so a change can be shown to be an improvement rather than assumed to be one.
AI systems fail in ways ordinary tests do not catch. The output is not wrong in a way an assertion detects; it is subtly worse, or right on the examples someone tried and wrong on the ones they did not. An eval suite is the answer: a fixed set of inputs, an expected outcome or grading rule for each, and a score that moves when the system changes.
The hard part is the grading. Some outputs can be checked exactly, such as whether the right record was updated. Most cannot, which is where a second model grades against a written rubric, or a human rates a sample. Neither is perfect, and both are far better than shipping on the strength of a handful of prompts that happened to look good.