A 30-question structured evaluation set covering five categories—facts, processes, code, comparisons, and edge cases—is Juejin author Chen Yingbo's "exam paper" sample for enterprise AI projects. Core judgment: evaluation sets (the question banks used to grade AI) aren't checklists—they're test assets with fields, categories, difficulty levels, and edge cases, determining what you measure and what regressions you miss.
What This Is
In Part 2 of the AI Application Production Handbook, Chen Yingbo flags three common pitfalls: ten ad-hoc questions can't represent real users; coverage capabilities go unexplained; the same answer scores 4 yesterday, 3 today—score drift. The fix is designing evaluation sets as test assets: diversity, layered difficulty, edge cases (deliberately picking questions likely to fail), and annotability (each question ships with an expected answer and scoring rubric).
The companion open-source repo ships a 30-question draft: 8 fact, 4 process, 4 code, 4 comparison, 8 edge cases, 2 hard questions. Difficulty: 6 easy / 15 medium / 9 hard, deliberately pushing hard + edge to 57%—because evaluation's mission is to catch regressions, not flaunt accuracy.
Another key distinction: the "Golden Set" stays locked down, acting as a regression gate against drift; the "Evolution Set" expands with business feedback. Both are essential—change the Golden Set and your baseline drifts.
Industry View
Consensus on "evaluation must be structured" runs high. But a counter-view deserves airtime: some Silicon Valley teams lean toward "ship and measure"—pushing live to see real feedback, arguing that polished evaluation sets are just another form of "open-book exam," and failure modes always live outside test coverage.
Our judgment: the two paths don't conflict. Evaluation sets don't replace real feedback—they install insurance against "ship and crash." The piece itself flags real pitfalls: evaluation set leakage (questions slipping into prompts or knowledge bases), survivorship bias (keeping only questions the AI answers well), and edge-case scorers too dumb to distinguish "correct refusal" from "no answer." Evaluation infrastructure is an engineering discipline, not a seat-of-the-pants exercise.
Impact on Regular People
For enterprise IT: When selecting AI vendors, ask one more question—do they have structured evaluation sets and regression reports? Demos aren't enough; you need real-scenario pass rates.
For individual careers: When using AI to draft plans or content, vendors that hand you a quantified scorecard will earn more trust than a vague "looks fine"—transparency becomes the new selling point for AI tools.
For consumer markets: "QA transparency" for AI products will gradually standardize, with evaluation pass rates mattering more than demos—a low-key shift already underway on the consumer side.