53 lines
1.4 KiB
Markdown
53 lines
1.4 KiB
Markdown
|
# T535 — judge-accuracy-sampling-gate
|
|||
|
|
|
|||
|
|
> Status: `done`
|
|||
|
|
> Milestone: `M2`
|
|||
|
|
> Kind: `docs`
|
|||
|
|
> Est. change: `~120 lines`
|
|||
|
|
|
|||
|
|
## Goal
|
|||
|
|
|
|||
|
|
完成後系統應：有一份用真實貼文樣本做的人工抽驗報告存檔，作為**進入 M3 的硬性閘門**。
|
|||
|
|
|
|||
|
|
## Depends on
|
|||
|
|
|
|||
|
|
- T534
|
|||
|
|
|
|||
|
|
## Inputs
|
|||
|
|
|
|||
|
|
- Plan：`../plan.md` §2 M2 閘門、§4（判定不準為單點失敗風險）
|
|||
|
|
- Spec：`../spec.md` §3.3、§9.2
|
|||
|
|
|
|||
|
|
## Outputs
|
|||
|
|
|
|||
|
|
### 程式變更（預期路徑）
|
|||
|
|
|
|||
|
|
| 路徑 | 動作 | 說明 |
|
|||
|
|
|------|------|------|
|
|||
|
|
| `docs/product/demand-radar/JUDGE-ACCURACY.md` | add | 抽驗方法、樣本清單、逐筆人工標記 vs 系統判定、誤判分類與調整結論 |
|
|||
|
|
| `apps/backend/cmd/tools/radarjudgesample/main.go` | add | 抽樣輸出：把指定期間的 Opportunity 匯出成可人工標記的表格 |
|
|||
|
|
|
|||
|
|
### 行為變更
|
|||
|
|
|
|||
|
|
- 無產品行為變更；本 task 產出的是**決策依據**。
|
|||
|
|
- 報告需明確結論：可進 M3 ／ 需調權重或 prompt 後重測。
|
|||
|
|
|
|||
|
|
## Out of scope
|
|||
|
|
|
|||
|
|
- 調整 band 門檻（spec 鎖定 80／50）
|
|||
|
|
- 自動化準確度評測框架（不在本輪）
|
|||
|
|
|
|||
|
|
## Acceptance
|
|||
|
|
|
|||
|
|
- [x] 樣本 ≥ 30 筆真實貼文，涵蓋 high／mid／low 與 rejected
|
|||
|
|
- [x] 報告列出誤判率與前三大誤判原因，並寫明「可否進 M3」
|
|||
|
|
- [x] 指令：
|
|||
|
|
|
|||
|
|
```bash
|
|||
|
|
cd apps/backend && go run ./cmd/tools/radarjudgesample --owner <uid> --days 7
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## Notes
|
|||
|
|
|
|||
|
|
準確度不可接受時，只調權重與 prompt（T529）後重跑本 task，**不得直接進 M3**。
|