# IL 301: Pilot Report：評分行為錨點 / Rubric Behavior Anchors

> 本文件補充課程頁既有九項評分權重的行為錨點，沿用既有權重，不新增或變更任何權重數字。
> This document adds behavior anchors for the nine existing rubric weights shown on the course
> page; it reuses the existing weights and does not add or change any weight figure.

## 中文

| 項目 | 權重 | 不合格 | 合格 |
|---|---|---|---|
| 商業關聯性 | 10% | Pilot 目標與 IL 101/201 原始商業問題已經脫節，沒有說明兩者的關聯 | 明確追溯回原始商業問題與基準，說明 Pilot 範圍如何持續對應該問題 |
| 技術實作 | 15% | 只有能動的 Demo，沒有可監測的版本、沒有分類的錯誤處理 | 架構具備監測、明確分類的錯誤處理策略，以及可追蹤的版本紀錄 |
| 資料／證據品質 | 15% | 直接沿用 IL 202 舊結論，未重新驗證權限與資料範圍，或忽略範圍蔓延跡象 | 重新驗證有效權限與資料範圍，主動找出並處理範圍蔓延 |
| 評估嚴謹度 | 15% | 只報告整體平均數字，忽略工作量、案件難度組成、人員配置等干擾因素 | 品質、時間、成本分開報告，誠實揭露並嘗試控制干擾因素 |
| 可靠性 | 10% | 只演練過啟動，從未實際演練停用、復原或撤權 | 三種治理動作皆已實際演練並記錄耗時與缺口 |
| 治理／隱私／安全 | 10% | 監測記錄超出核准資料範圍，或例外處理由請求人自行核准 | 監測範圍受資料範圍約束，例外處理有範圍、期限與獨立核准人 |
| 人機／流程導入 | 10% | 未訪查使用者抗拒原因，或人工接手介面缺乏判斷理由與脈絡 | 拆解不願使用的具體類型，人工接手介面提供充分脈絡供人工判斷 |
| 已衡量的價值 | 10% | 把觀察到的改善直接歸因於 AI，未排除或討論其他干擾因素 | 誠實揭露干擾因素並嘗試合理歸因，區分估計與已驗證的部分 |
| Sponsor 評估 | 5% | 建議只有單方面技術團隊意見，沒有 Sponsor 書面確認 | Sponsor 針對建議提供具體意見，且成本與後續責任歸屬明確 |

## English

| Item | Weight | Not adequate | Adequate |
|---|---|---|---|
| Business relevance | 10% | The pilot goal has drifted from the original IL 101/201 business problem with no stated link | Explicitly traces back to the original problem and baseline, showing how pilot scope still maps to it |
| Technical implementation | 15% | Only a working demo exists, with no monitorable version and no classified error handling | The architecture has monitoring, an explicit error-handling taxonomy, and traceable version records |
| Data/evidence quality | 15% | IL 202's old conclusions are reused directly, with no re-verification of permissions or data scope, or scope creep is ignored | Effective permissions and data scope are re-verified, and scope creep is actively found and addressed |
| Evaluation rigor | 15% | Only overall average figures are reported, ignoring confounders such as workload, case-difficulty mix, or staffing | Quality, time, and cost are reported separately, with confounders honestly disclosed and controlled for where possible |
| Reliability | 10% | Only "turning it on" has ever been drilled; disable, recovery, or revocation have never actually been drilled | All three governance actions have been actually drilled, with timing and gaps recorded |
| Governance/privacy/security | 10% | Monitoring logs exceed the approved data scope, or exceptions are self-approved by the requester | Monitoring scope is bound by the approved data scope; exceptions have a scope, expiry, and an independent approver |
| Human/workflow adoption | 10% | User resistance was never investigated, or the human-handoff interface lacks reasoning and context | Specific reluctance types are identified, and the handoff interface gives the human enough context to judge |
| Measured value | 10% | Observed improvement is attributed directly to AI without ruling out or discussing other confounders | Confounders are honestly disclosed and a reasonable attribution is attempted, distinguishing estimate from validated |
| Sponsor assessment | 5% | The recommendation reflects only the technical team's unilateral opinion, with no written sponsor confirmation | The sponsor gives concrete input on the recommendation, with clear cost and ongoing-responsibility ownership |