GDPval with NOOA

The NOOA recipe uses GDPval’s existing environment server, collection driver, resources server, and grading configuration. It selects the harness-neutral prompt and runs NOOA inside the resources-owned sandbox.

Use GDPval’s existing prerequisites: an image built from the GDPval Dockerfile, GDPVAL_SANDBOX_IMAGE, OPENSANDBOX_DOMAIN, OPENSANDBOX_API_KEY, TAVILY_API_KEY, and credentials for the configured judge endpoints. Set HF_TOKEN when needed for reference-file downloads. Set PERSIST_DELIVERABLES_DIR to a writable directory for collected submissions. See the canonical GDPval instructions for judge-panel configuration and re-judging.

Put your policy endpoint and deployment overrides in a private private-policy-provider.yaml overlay. Ensure the sandbox can reach the Gym model and resources servers; use ++use_absolute_ip=true when the default addresses are unreachable. From the repository root:

gym eval prepare --benchmark gdpval/nooa
gym env resolve \
    --config benchmarks/gdpval/nooa.yaml \
    --config responses_api_models/vllm_model/configs/vllm_model.yaml \
    --config private-policy-provider.yaml
gym eval run \
    --config benchmarks/gdpval/nooa.yaml \
    --config responses_api_models/vllm_model/configs/vllm_model.yaml \
    --config private-policy-provider.yaml \
    --split benchmark --limit 1 --concurrency 1 --num-repeats 1 \
    --output results/gdpval-nooa.jsonl

The existing direct task tools are available to NOOA: finish(reason, paths), abandon_task_finish, web_search, and fetch_web_page. finish validates and records a submission but does not terminate NOOA; the agent must still return its native TaskResult. GDPval’s existing missing-finish fallback and grading behavior remain in effect.