math_with_judge:
  resources_servers:
    math_with_judge:
      entrypoint: app.py
      judge_model_server:
        type: responses_api_models
        name: policy_model
      judge_responses_create_params: {input: []}
      should_use_judge: false
      domain: math
      verified: false
      description: OpenClaw agent harness on OpenMathReasoning math dataset with math-verify
      value: Improve model math capabilities in the OpenClaw agent harness

math_with_judge_openclaw_agent:
  responses_api_agents:
    openclaw_agent:
      entrypoint: app.py
      resources_server:
        type: resources_servers
        name: math_with_judge
      concurrency: 8
      command: openclaw
      model: nvinf/nvidia/qwen/qwen3-next-80b-a3b-instruct
      openclaw_version: 2026.6.10
      thinking: "off"
      timeout: 600
      system_prompt: |
        You are a precise math assistant with an exec/terminal tool that runs shell commands.
        You MUST verify every numeric result by running Python via your exec tool (use `python3 -c "..."`,
        not `python`) before answering. Do not compute large values by hand. After the tool returns,
        state your final answer as \boxed{answer}.
      openclaw_config:
        models:
          providers:
            nvinf:
              api: openai-completions
              baseUrl: ${policy_base_url}
              apiKey: ${policy_api_key}
              models:
              - id: nvidia/qwen/qwen3-next-80b-a3b-instruct
                name: nvidia/qwen/qwen3-next-80b-a3b-instruct
                api: openai-completions
      datasets:
      - name: example
        type: example
        jsonl_fpath: resources_servers/math_with_judge/data/example.jsonl