Sundowner

after_step 决定触发时机，observation_refs 限定证据，choices 绑定允许执行的步骤。

{
  "id": "route",
  "after_step": "ask",
  "observation_refs": [
    "first-reply"
  ],
  "driver_purpose": "decision",
  "decision_schema": "choice@1",
  "condition": "先调用只读观察工具检查当前回复。如果被测对象明确回复 SD_READY_1，选择 continue；否则选择 clarify。只返回已声明选项。",
  "scope": {
    "resource_refs": [
      "room"
    ],
    "step_refs": [
      "continue",
      "clarify"
    ]
  },
  "choices": [
    {
      "id": "continue",
      "step_ids": [
        "continue"
      ]
    },
    {
      "id": "clarify",
      "step_ids": [
        "clarify"
      ]
    }
  ],
  "limits": {
    "max_decisions": 1,
    "timeout_ms": 120000,
    "max_output_bytes": 12000
  },
  "on_failure": "block_case"
}

动态模块补齐 run、node、attempt 和 point 标识。执行器只接受已声明的选择；未声明生成能力时不能随意改写步骤。

{
  "run_id": "run-model-smoke",
  "node_id": "two-round/dynamic/route",
  "attempt_id": "attempt-1",
  "point_id": "two-round/route",
  "status": "selected",
  "choice_id": "continue",
  "evidence_ids": [
    "reply-1"
  ],
  "rationale": "已读取首轮回复，内容为 SD_READY_1，满足继续条件。"
}

声明用途、模型、指令、可用只读工具和轮数预算。

{
  "purpose": "dynamic",
  "model_ref": "example-model",
  "instructions": "You are the dynamic decision maker of an evaluation harness. At a pre-declared decision point you select exactly one of the declared choices, using only the observations the host shares with you in this session.\nYou do not know and must not guess hidden truths, later plan steps, scoring or anything outside the shared observations. If the shared evidence does not satisfy the condition, pick the choice that the condition text designates for that case.\nBefore citing an observation, read it in full with read_observation. Cite only observation ids you have read. Use list_observations / read_coverage / read_execution_context when you need the current state.\nFinal answer: status \"output\" with output matching output_schema: {\"choice_id\": <declared id>, \"evidence_ids\": [<observation ids you read and rely on>], \"rationale\": <= 400 characters, \"generated_text\": only when generation is declared and your choice includes it — obey max_chars and forbidden_patterns; omit the field otherwise}.\nNever ask for tools, files or resources that are not declared.",
  "context_refs": [],
  "tools": [
    {
      "name": "read_observation",
      "effect": "read",
      "input_schema": {
        "type": "object",
        "additionalProperties": false,
        "required": [
          "id"
        ],
        "properties": {
          "id": {
            "type": "string",
            "minLength": 1
          }
        }
      }
    }
  ],
  "limits": {
    "max_turns": 2,
    "timeout_ms": 120000,
    "max_tokens": 8000
  }
}

模型可见输入包含条件、观察索引、完整性和当前执行状态，不包含隐藏真值或未来步骤内容。

{
  "session_ref": "example-session",
  "input": {
    "decision": {
      "purpose": "decision",
      "condition": "先调用只读观察工具检查当前回复。如果被测对象明确回复 SD_READY_1，选择 continue；否则选择 clarify。只返回已声明选项。",
      "reference_text": null,
      "choices": [
        {
          "id": "continue"
        },
        {
          "id": "clarify"
        }
      ],
      "generation": null
    },
    "observations": [
      {
        "id": "reply-1",
        "kind": "message",
        "occurred_at": "2026-09-10T00:00:00.200Z",
        "subject_ref": "assistant",
        "resource_ref": "room",
        "preview": "…"
      }
    ],
    "coverage": [
      {
        "id": "coverage-1",
        "requirement_id": "first-reply",
        "kind": "message",
        "status": "complete",
        "reason": null,
        "from": "2026-09-10T00:00:00.000Z",
        "to": "2026-09-10T00:00:01.000Z"
      }
    ],
    "execution_context": {
      "completed_cases": [],
      "not_executed_cases": [],
      "executed_steps": [
        "two-round/ask"
      ],
      "selected_choices": [],
      "active_fact_ids": [],
      "active_branches": []
    },
    "output_schema": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "choice_id",
        "evidence_ids",
        "rationale"
      ],
      "properties": {
        "choice_id": {
          "type": "string",
          "enum": [
            "continue",
            "clarify"
          ]
        },
        "evidence_ids": {
          "type": "array",
          "items": {
            "type": "string"
          },
          "uniqueItems": true
        },
        "rationale": {
          "type": "string",
          "maxLength": 400
        }
      }
    },
    "note": "Read the observations you rely on with read_observation, then answer with status \"output\"."
  },
  "tool_results": []
}

模型返回工具请求，宿主执行允许的只读工具，并通过下一次 turn 传回结果。

{
  "status": "ok",
  "value": {
    "session_ref": "example-session",
    "status": "tool_requests",
    "output": null,
    "tool_requests": [
      {
        "id": "read-1",
        "name": "read_observation",
        "input": {
          "id": "reply-1"
        }
      }
    ],
    "usage": {
      "input_tokens": 0,
      "output_tokens": 0
    }
  }
}

判官输出 pass、fail、indeterminate 或 not_applicable，并关联具体证据。算分脚本汇总这些结果。

{
  "case_id": "two-round",
  "criterion_id": "ready",
  "status": "pass",
  "basis": "evidence",
  "score": 1,
  "evidence_ids": [
    "reply-1"
  ],
  "coverage_refs": [
    "coverage-1"
  ],
  "reason": "首轮回复满足明确回复 SD_READY_1 的判据。",
  "judge": {
    "id": "semantic-response",
    "version": "1",
    "model_ref": "example-model"
  }
}