Eval strategy
Evals are the unit tests of AI systems. Start from real failure modes, write targeted checks and run them on every prompt, model or tool change.
Stable: this rarely changes. Reviewed 17 Oct 2026.
Key ideas
- Do error analysis first: read 50 to 100 real traces, label failures, group them into categories.
- Write evals per failure category, not generic "quality" scores.
- Layers: code-based assertions (format, required fields), LLM-as-judge for fuzzy criteria, human review for calibration.
- Keep an offline regression suite in CI and sample production traffic for online evals.
- Track pass rates per category over time and per model version.
Common mistakes
- Off-the-shelf metrics like helpfulness scores that do not map to real failures.