给既有回答打分，别把自报信心当已校准概率
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
| question | assistantAnswer | authorCorrectness | sourceFacts |
|---|---|---|---|
| 活动哪天？ | 2026-11-07 | correct | |
| 费用多少？ | ¥88 | incorrect | |
| 停车免费吗？ | 可以免费停车。 | unsupported | |
| {"date": "2026-11-07", "fee": "¥68", "cancel": "提前24小时"} |
选择匹配adapter
reference.csv：作者验收映射，不是目标API回包
| scoreSource | sampleCount | status | answerVersion |
|---|---|---|---|
| uncertainty adapter actual output only | |||
| 3 | protocol demonstration, not calibration proof | ||
| must remain fixed after scoring |
应当保留的约束
| scoreSource | sampleCount | status | answerVersion |
|---|---|---|---|
| uncertainty adapter actual output only | |||
| 3 | protocol demonstration, not calibration proof | ||
| must remain fixed after scoring |
故意构造的错误样例
| scoreSource | error | sampleCount | claim |
|---|---|---|---|
| chat text I am99% sure | not intrinsic output | ||
| unsupported generalization | 3 | all-domain calibrated |
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。