Shipping LLM-powered features without a test strategy leads to flaky CI and silent quality regressions. Traditional assertions fail because the same prompt can produce valid variations. Evaluation harnesses replace brittle equals checks with structured scoring.
Priya Sharma
QA Architect · 20 February 2024
Shipping LLM-powered features without a test strategy leads to flaky CI and silent quality regressions. Traditional assertions fail because the same prompt can produce valid variations. Evaluation harnesses replace brittle equals checks with structured scoring.
01
Curate representative prompts and expected properties—not necessarily exact wording.
02
Score completeness, safety, and tone separately so failures are actionable.
03
Log structured metadata and review a random sample weekly for drift.
import { evaluateResponse } from '@adt/llm-eval';
export async function assertSupportReply(output, context) {
const result = await evaluateResponse({
output,
rubric: [
'Answers the billing question directly',
'Does not invent refund policies',
'Tone is professional and concise',
],
context,
minScore: 0.85,
});
expect(result.passed, result.failures.join('; ')).toBe(true);
}Check type | Example | Failure handling |
|---|---|---|
Guardrail | No medical advice | Block response, alert on-call |
Quality | Answers billing question | Warn, route to human review |
Style | Professional tone | Log only in lower environments |
“Flaky LLM tests usually mean your rubric is vague—not that the model is random.”
| Context | A support copilot drafts replies from knowledge-base articles. Replies must be accurate and on-brand. |
|---|---|
| Challenge | String equality fails on paraphrases; manual review does not scale to every PR. |
| Solution | CI runs fifty golden prompts through the copilot; rubric scorer enforces minimum scores per dimension. Guardrails block policy violations outright. |
| Outcome | PR feedback time drops; two policy regressions are caught in CI before merge. |
Evaluation does not stop at merge. Sample live responses, compare rubric scores week over week, and alert when guardrail violations spike. Pair automated scoring with a lightweight human review queue for scores near the threshold.
LLM features are testable when you redefine correctness as bounded semantic quality plus hard guardrails. Human reviewers still approve rubrics and golden sets—the same way they approve test plans for traditional features.