Gold Set · AI.WIKI

01 · 十道重复的简单题，能替代一道真正危险的边界题吗01 · Can ten duplicate easy cases replace one critical boundary case?

题库从 6 条涨到 12 条，但新增的全是近似副本，退款边界仍然空白——你对 92% 这个分数会更有信心吗？The set grows from six cases to twelve, but every new case is a near-duplicate and the refund boundary stays empty. Should a 92% score feel more certain?

02增删边界案例，观察覆盖矩阵、置信区间与重复/泄漏标记联动Add or Remove Boundary Cases and Watch Coverage, Confidence, Duplicates, and Leakage Move Together

评测证据血缘 · 蓝/黄/红eval lineage · blue/yellow/red

状态 AMetric A

—

状态 BMetric B

—

状态 CMetric C

—

所有概率、分数、成本、token 与输出均为固定种子的教学模拟，不代表真实模型输出。把主变量拖回即可复核结论。All probabilities, scores, costs, tokens, and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.

03把刚才的变化拆成 4 个零件Break the Change into Four Parts

Gold Set

RUBRIC

审定量尺Reviewed rubric

每条样本都有期望结果或明确评价规则，争议项先复核。Every case has an expected result or explicit rubric; disputed cases are reviewed first.

COVER

边界覆盖Boundary coverage

正常、缺字段、歧义与越权四类格子都需要真实样本。Normal, missing-field, ambiguous, and unauthorized cells all need real cases.

CLEAN

独立与去重Independence and dedupe

近似重复不会增加等量证据；进入提示或知识库的样本必须标泄漏。Near-duplicates do not add equal evidence; cases seen by prompts or retrieval must be marked leaked.

UNCERT

不确定性Uncertainty

样本少时区间宽；覆盖增长且数据独立时，估计才逐渐收紧。Intervals remain wide with few cases and tighten only as independent coverage grows.

04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction

MECHANISM

CI ≈ p̂ ± 1.96 √(p̂(1 − p̂) / n_clean)

主交互用干净且非重复的有效样本数 n_clean 计算教学区间；加重复样本不会假装让区间变窄。The interaction uses only clean, non-duplicate cases for the teaching interval. Duplicate cases do not pretend to narrow uncertainty.

加边界样本Add boundary case覆盖空格fill a coverage gap

自动扫描Auto-scan重复 / 泄漏亮标flag duplicate / leaked

重算证据Recompute evidence覆盖 + 区间coverage + interval

05连续加入两个重复样本，再把一条样本标为泄漏Add Two Duplicate Cases, Then Mark One Case as Leaked

主动制造失败Create a failure

边界实验Boundary experiment

总样本数会增加，但干净独立证据不会同比增长。实验应自动亮红并拒绝把这些样本当成新增置信。The raw case count rises, but clean independent evidence does not. The lab must flag the cases and refuse to treat them as new confidence.

尚未执行。先在主交互中观察正常机制。Not run yet. Observe the normal mechanism in the main interaction first.

✗ 黄金集就是尽可能大的题库。A gold set is simply the largest possible set.→ ✓ 代表性、审定量尺与独立性比原始条数更关键。Representativeness, reviewed rubrics, and independence matter more than raw count.

✗ 重复样本也能等量缩小置信区间。Duplicates narrow the interval equally.→ ✓ 近似副本不提供等量独立信息。Near-copies do not provide equal independent information.

✗ 样本进过提示库仍可当干净黄金题。A case remains clean after entering the prompt library.→ ✓ 血缘一旦相交就要标泄漏并隔离该证据。Once lineage intersects, mark leakage and isolate that evidence.

06 · 一句话带走06 · One line to keep

黄金集的价值来自审定量尺、关键边界覆盖和独立干净样本；重复或泄漏只会让表面题数膨胀，不会增加可信证据。A gold set earns trust through reviewed rubrics, critical boundary coverage, and clean independent cases. Duplicates and leakage inflate counts, not evidence.