Evals And Observability hard
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?
- A Raise the failure threshold to a 10-point drop so the gate stops tripping on noise, and accept that it will catch fewer real regressions.
- B Remove the offline gate and rely on online thumbs-up rates, because live user feedback is the actual ground truth for prompt quality.
- C Measure run-to-run variance, use repeated samples per case with a paired test, and set the threshold outside the noise band.
- D Force temperature 0 on both the generator and the judge so reruns are fully deterministic, then keep the 3-point threshold as it is.
Answer: C. Measure run-to-run variance, use repeated samples per case with a paired test, and set the threshold outside the noise band.
A gate whose threshold sits inside its own noise band trains engineers to ignore it. **Repeated samples and a paired comparison** shrink the noise band so a small threshold becomes meaningful, at the cost of more eval tokens and CI time per change. Forcing temperature 0 (D) is tempting, but it evaluates a different configuration from the 0.7 used in production, and serving nondeterminism means runs are still not fully reproducible. It is only reasonable when production also runs greedy decoding.
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 RAG assistant's answer-quality score dropped from 82 to 70 after a change. You only measure final answer quality. Why is that a problem for debugging?