Evals And Observability hard
You are choosing between two models for extracting fields from legal contracts. Model X scores 6 points higher on a widely cited public reasoning benchmark. Model Y costs about one third as much per token at list price. How should you decide?
- A Choose X, since a general benchmark lead of 6 points reliably transfers to domain extraction quality and justifies the higher cost.
- B Choose Y, since public benchmarks carry no signal at all for production work and per-token cost is the only real differentiator.
- C Run both on a task-specific golden set of 100 to 200 labeled contracts with field-level metrics, then compare quality per unit cost.
- D Run both on five representative contracts in a live demo with the legal team, since qualitative review catches nuance that metrics miss.
Answer: C. Run both on a task-specific golden set of 100 to 200 labeled contracts with field-level metrics, then compare quality per unit cost.
Public benchmark accuracy measures general capability, not your task, so the ranking often flips on domain extraction. A **task-specific eval** of 100 to 200 cases costs labeling effort but measures the real quality-versus-cost tradeoff. A five-contract demo (D) feels grounded, but the variance across so few cases is too high to separate two strong models. Benchmarks are a reasonable first filter only when you have no labeled task data yet.
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?