分清yes/no原始输出与Mellea连续分数，阈值由项目设定
本页展示随包原生输入和验收参考。目标产品运行状态：未运行。表格与图片不代表产品已经生成的结果。
input.csv / practice-data.json
本课实际输入字段
| candidate | text | requirements | interface | actualScore | threshold |
|---|---|---|---|---|---|
| good | 2026-11-07到木作室参加体验，费用¥68。 | ["日期2026-11-07", "地点木作室", "费用¥68", "不得承诺免费"] | |||
| bad | 欢迎到木作室免费体验。 | ["日期2026-11-07", "地点木作室", "费用¥68", "不得承诺免费"] | |||
| Mellea | None | project-defined after calibration |
核对adapter配对
reference.csv：作者验收映射，不是目标API回包
| candidate | authorExpected | adapterOutput | melleaOutput |
|---|---|---|---|
| good | satisfies stated requirements | ||
| bad | fails date/price/no-free requirements | ||
| {"score":"yes"} or {"score":"no"} | float0..1 |
应当保留的约束
| candidate | authorExpected | adapterOutput | melleaOutput |
|---|---|---|---|
| good | satisfies stated requirements | ||
| bad | fails date/price/no-free requirements | ||
| {"score":"yes"} or {"score":"no"} | float0..1 |
故意构造的错误样例
| test | claim | error | base | adapter |
|---|---|---|---|---|
| bool(0.2) | requirements passed | truthiness is not threshold | ||
| size mismatch | granite-4.1-3b | requirement-check/granite-4.1-8b/lora |
先保存真实产品输出，再逐项对照。actualOutput / actualResult 在未执行时保持 not_run，不能把这里的期望值复制成已测结果。