多模型作文評分

osop_version: "1.0"
id: "multi-model-grading"
name:"多模型作文評分"
description:"三個 LLM 並行評分，彙整分數，進行偏見檢查，再發出最終成績。"

nodes:
  - id: "essay_intake"
    type: "system"
    name: "作文接收"
    description: "接收提交的作文及評分標準，對學生身份進行匿名化處理。"

  - id: "grader_1"
    type: "agent"
    subtype: "llm"
    name: "評分者 A（GPT-4o）"
    description: "依結構、論點品質、證據運用與寫作清晰度評分。"

  - id: "grader_2"
    type: "agent"
    subtype: "llm"
    name: "評分者 B（Claude）"
    description: "使用相同評分標準獨立評分，對其他評分者保持盲評。"

  - id: "grader_3"
    type: "agent"
    subtype: "llm"
    name: "評分者 C（Gemini）"
    description: "作為第三位評分者獨立評分，以形成穩健的共識。"

  - id: "aggregate"
    type: "system"
    name: "分數彙整"
    description: "計算加權平均分數，若任一評分者偏差超過平均值 15% 則標記。"

  - id: "bias_check"
    type: "agent"
    subtype: "llm"
    name: "偏見偵測代理人"
    description: "分析評分模式是否存在人口統計偏見、主題偏見或長度偏見。"

  - id: "final_grade"
    type: "agent"
    subtype: "llm"
    name: "最終成績與回饋"
    description: "發出最終成績，附上所有評分者的綜合回饋與改善建議。"

edges:
  - from: "essay_intake"
    to: "grader_1"
    mode: "parallel"
  - from: "essay_intake"
    to: "grader_2"
    mode: "parallel"
  - from: "essay_intake"
    to: "grader_3"
    mode: "parallel"
  - from: "grader_1"
    to: "aggregate"
    mode: "parallel"
  - from: "grader_2"
    to: "aggregate"
    mode: "parallel"
  - from: "grader_3"
    to: "aggregate"
    mode: "parallel"
  - from: "aggregate"
    to: "bias_check"
    mode: "sequential"
  - from: "bias_check"
    to: "final_grade"
    mode: "conditional"
    when: "bias.detected == false"
  - from: "bias_check"
    to: "grader_1"
    mode: "fallback"
    label: "Bias detected, re-grade with adjusted prompts"