Evals And Observability medium
Your CI runs a 150-case offline eval on every prompt change, scored on a 100-point rubric. Re-running the same prompt shows scores varying by about about 1.5 points. How should you set the regression gate?
- A Fail on a drop of roughly 2 to 5 points against a pinned baseline, chosen to sit just above the eval's measured run-to-run noise.
- B Fail on any drop at all, since even a 0.5-point decrease is a real regression that will compound across future prompt changes.
- C Fail only on drops of 15 points or more, since smaller differences on 150 cases are noise and a gate should rarely block shipping.
- D Skip a numeric threshold and require manual review of a few sample outputs, since rubric scores miss regressions that humans notice.
Answer: A. Fail on a drop of roughly 2 to 5 points against a pinned baseline, chosen to sit just above the eval's measured run-to-run noise.
A gate set slightly above the measured noise, commonly a **2 to 5 point drop** on a 100-point rubric, catches real regressions without blocking on randomness. The tradeoff is between false alarms and missed regressions. Failing on any drop (B) turns about 1.5 points of noise into constant false failures that teams learn to override. A 15-point threshold (C) would only make sense for a tiny, very noisy set where nothing smaller is detectable.
One system design question a day, with spaced repetition so the ones you miss come back.
Get the appRelated questions
- In one release, a RAG assistant got a new embedding model and a new chunking strategy, and the end-to-end judge score dropped 6 points. The team only tracks the end-to-end score and has spent a week tweaking the generation prompt with no recovery. What is the best next step?
- Over two months, your team iterated through 40 prompt versions, each tuned against the same 150-case golden set. The offline score rose from 72 to 91. Online task completion and retry rate have not changed. What is the most likely problem, and what is the fix?
- A prompt change produced an impressive answer in a live demo for leadership. The PM wants to ship it today. What should you do?
- You plan to replace human graders with an LLM judge for your answer-quality eval. Before trusting its scores in a CI gate, what should you do first?
- After you swap to a smaller model, p95 latency falls from 900 ms to 350 ms and the ops dashboard is all green. Leadership asks whether answer quality is still fine. What is the right response?
- Your CI gate fails any prompt change that drops the score 3 or more points on a 100-point rubric over 60 cases. Production generates at temperature 0.7. Running the eval twice on the same commit gives scores that differ by up to 4 points, and engineers have started rerunning failed gates until they pass. How do you fix the gate?