Hosted Arena evaluation

Hosted Arena submissions use the existing private GPU broker and its pinned, isolated worker image. They share FIFO scheduling, the global compute semaphore, session/IP/service run budgets, and workspace expiry with ordinary Studio jobs. An Arena evaluation trains a backbone at up to four parameter budgets, so it has a 7-day wall-clock guard; ordinary Studio jobs retain their 30-minute limit. This operational allowance does not change the frozen 240 complete epochs and 94,320 optimizer updates per seed and configuration on APA_200MHz_b, the data, seeds, validation policy, or scoring gates. The API VM does not run Arena training.

Enable submissions with OPENDPD_WEB_ARENA_SUBMISSIONS=1 on the web service, or WebConfig(arena_submissions=True) when embedding it. The default is disabled. The service must also have its existing private GPU token configured. Build and pin a worker image containing the same OpenDPD version, Arena runner, protocol, and packaged calibration assets as the API service. Publishing an image or restarting a service is an operator action; setting the flag alone does not perform either action.

The Arena allowance is bounded by WebConfig.arena_max_runtime_seconds (604,800 seconds by default and at most); reducing it does not change the ordinary-job setting. Local Studio and the official benchmark launcher use the same 7-day Arena guard. A running worker's original start remains the deadline origin when an operator hands it to a replacement supervisor; supervision does not restart training or add epochs.

At startup the host agent probes Arena support inside that pinned image and verifies the packaged data and PA weight hashes. It advertises the protocol hash over the existing private broker connection. The submission API and Studio's Submit view become available only while the live agent advertises the exact protocol accepted by the API. An older image, missing assets, a stale heartbeat, or a different protocol keeps submission unavailable while rankings remain readable.

Each required condition and seed consumes one unit of the existing run quota, and that unit trains the case at every budget of the parameter sweep. Thus a neural entry uses three units on the APA_200MHz_b board, and a deterministic baseline uses one unit. Each entry trains up to four models per unit and can occupy the shared worker for up to the 7-day guard. Operators who enable hosted submissions should size runs_per_session, runs_per_ip and runs_per_day with that in mind. A request for a stale protocol or an unknown backbone is refused before any budget is charged. Both ordinary runs and Arena entries count toward the same pending-job limits; opening another session does not reset the IP or service budget.

The broker sends only the server-validated request and optional bounded template definition. Calibration data and evaluators come from the read-only image. Container output is limited to the result, progress, and worker log; it cannot overwrite the authoritative request or submission. The API checks the returned model identity, re-derives the registered sweep and recomputes the operation counts, scores, required case coverage, frozen PA hashes, and the output-power gate before displaying a rank. No host is timed, so a hosted result is directly comparable with the shipped rows. Closing or expiring the workspace cancels queued and active evaluations, and late results cannot restore them.

Results remain private to the temporary workspace and disappear with its normal cleanup. A hosted submission does not publish a global leaderboard record.

Isolated container smoke test

Hosted submissions are off by default. Before enabling them on a deployment, run one real four-configuration MP submission through the pinned GPU image and confirm that it completes: the host agent unpacks the broker's input, the unprivileged, network-isolated container writes its result, and only the selected output files are transferred back. At startup the agent already runs the pinned image's exact-protocol capability check and a cold CUDA probe. Automated tests cover the agent and broker transfer contract with the container stubbed (tests/unit/test_arena_gpu_adapter.py); they do not start a container or use a GPU.

The v6 local and hosted runners prefer CUDA when available, including zero-threshold Delta cells with reviewed dense training equivalents. The distributed reference launcher may assign selected fits to CPU and records the actual host/device. TF32 is disabled. These execution paths have dedicated forward, gradient, optimizer and inference checks. The container smoke checks the execution boundary; the complete 218-fit reference has a separate independent audit.