AI features without evals are live theater
AI can demo perfectly and fail on real distributions. Pre-ship eval sets make acceptable behavior measurable: correctness, refusals, safety, format, latency, cost. We wire evals into CI so prompt/model changes re-run. Below is a minimal system two engineers can start. Fold this into KSX Studio discover→design→build→launch: named owner, Definition of Done, staging proof, and rollback so verbal alignment does not drift.
Seed from real failures
Support tickets, sales objections, internal complaints are best seeds. Each item: input, context, expected points, unacceptable behaviors. Fold this into KSX discover→design→build→launch checks on staging so verbal alignment does not drop the requirement. Fold this into KSX Studio discover→design→build→launch: named owner, Definition of Done, staging proof, and rollback so verbal alignment does not drift.
Dimensions—not one score
Score correctness, completeness, citations, refusal, safety, format, toxicity separately. A single “good?” cannot regress. Fold this into KSX discover→design→build→launch checks on staging so verbal alignment does not drop the requirement. Fold this into KSX Studio discover→design→build→launch: named owner, Definition of Done, staging proof, and rollback so verbal alignment does not drift.
Gold labels and rater agreement
Dual-review disputed items; write rubrics. Mark ambiguous items rather than force labels. Fold this into KSX discover→design→build→launch checks on staging so verbal alignment does not drop the requirement. Fold this into KSX Studio discover→design→build→launch: named owner, Definition of Done, staging proof, and rollback so verbal alignment does not drift.
Automate what you can
Automate format, JSON schema, keyword/regex, retrieval hits; semantic quality may use model assist with human spot checks. Fold this into KSX discover→design→build→launch checks on staging so verbal alignment does not drop the requirement. Fold this into KSX Studio discover→design→build→launch: named owner, Definition of Done, staging proof, and rollback so verbal alignment does not drift.
Gate thresholds and release
Block release if core-set pass rate fails. Known failures allowed only with owner signature and expiry. Fold this into KSX discover→design→build→launch checks on staging so verbal alignment does not drop the requirement. Fold this into KSX Studio discover→design→build→launch: named owner, Definition of Done, staging proof, and rollback so verbal alignment does not drift.
Online drift feedback loop
Redact production bad cases into the eval set. Prompt-ops changes must attach eval diffs. Fold this into KSX discover→design→build→launch checks on staging so verbal alignment does not drop the requirement. Fold this into KSX Studio discover→design→build→launch: named owner, Definition of Done, staging proof, and rollback so verbal alignment does not drift.
Cost and latency are metrics
Log tokens and latency in evals so “much better but 10× cost” does not ship unconsciously. Fold this into KSX discover→design→build→launch checks on staging so verbal alignment does not drop the requirement. Fold this into KSX Studio discover→design→build→launch: named owner, Definition of Done, staging proof, and rollback so verbal alignment does not drift.
Checklist
- 1≥50 core cases including safety/refusal
- 2Rubric documented
- 3CI runs automated slice
- 4Release thresholds in README
- 5Weekly production bad-case intake
Key takeaways
- Eval sets are regression tests for AI products.
- Multi-dimension scores guide prompt and model changes.
- Gates and feedback loops remove luck from quality.
FAQ
- How many cases?
- Start 50–100 high-value; grow with features. Coverage beats blind volume.
- Fully automatic scoring?
- Partly; critical safety/correctness need human sampling early.
- Public benchmarks instead?
- Supplement only—not a substitute for your distribution and refusal policy.