Skip to content

[測試·P1] VY 鏡片判準對抗稽核:golden set + cwc audit.md 三層稽核 #120

Description

@atomchung

背景

Learning repo 2026-07-05 對讀 Anthropic 官方 cwc-workshops 時發現 rightmodel workshop 的 eval 稽核清單(216 行,框架無關),多項直接適用 VY 鏡片的判準品質稽核。鏡片是產品核心,判準品質=「eval 瓶頸是判準不是工具」的實戰場。

目標

對 VY 鏡片建立對抗稽核,讓「鏡片會不會誤判」有機械答案,而不是靠執行當下的 self-check 自律(#82 同根源)。

工作項(從 audit.md 挑適用的)

  1. Golden set Tier 1 程式化掃描:重複率/label 平衡/schema 破損。沒有 golden set 就先建 20–50 筆,含 probe 邊界案例(零 probe=只跑過 happy path)
  2. 對抗性 CSV 集:極端案例餵鏡片看誤判——拆股、多幣別([P2·邊界] 非美股/多幣別:台股港股 CSV 會靜默算錯 — 至少明示邊界 + 偵測警告 #51)、靜默丟資料([P2·誠實] 輸入層靜默丟資料:load() 無「讀入/跳過」計數,與卡面的數字誠實不一致 #50)等已知邊界+刻意構造的 wrong-but-plausible 案例
  3. per-task auditor:用 audit.md 附的 JSON prompt 模板逐題 flag:ambiguous / gold_suspect / grader_too_strict / grader_too_lenient / trivially_cheatable
  4. known-good / known-bad 煙霧測試:oracle 輸入該全過、null baseline 該全掛——不過就是判準壞了不是資料壞了
  5. spot-check failures:抽讀被鏡片標「有問題」的復盤案例,若 >1/10 其實是判準錯,先修判準再談覆蓋率
  6. 鏡片若含 LLM judge call:餵 known negatives(空輸入/自信答錯題)確認會 fail;rubric 要 concrete 可 check,不是 vibes

參考

來源:Learning repo Fable 視窗掃描優先序第 3 條(notes/fable-limited-window-strategy.md

🤖 Generated with Claude Code

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions