# Argus · Nebius execution record and cloud asset gate

Status on 2026-10-22: **Token Factory model inference completed; an AI Cloud H100 job processed all twelve fox source-bound photo/render pairs and produced verified diagnostic outputs.** The fox mesh, textures, and fixed-camera renders shown in the demo were computed locally. The cloud stage audited those saved assets; it did not reconstruct or promote a new model. See [the twelve-view diagnostic and cost report](ARGUS_NEBIUS_DIAGNOSTIC_REPORT_20261022.md).

## Verified execution

| Service | Receipt | What ran | What it establishes |
| --- | --- | --- | --- |
| Nebius Token Factory | `docs/assets/cloud/nebius-receipt.json` | One live call to NVIDIA `nvidia/nemotron-3-super-120b-a12b`, 779 prompt and 911 completion tokens | Nebius-hosted NVIDIA model inference for a bounded diagnostic recommendation; no authority to promote assets |
| AI Cloud Serverless Jobs, US North | `aijob-u02ffyc6wx2f2ykpps`, completed | `nvidia-smi` on B300 | Container startup and GPU visibility only |
| AI Cloud Serverless Jobs, US North | `aijob-u02pkp4rdsh485zwhv`, cancelled | B300 Argus replay attempt; provisioning retried twice and never started the container | No successful workload result |
| AI Cloud Serverless Jobs, EU North | `aijob-e00y9dmjj6bybqpy6r`, completed | CUDA/PyTorch arithmetic on twelve precomputed MAE pairs using H100 | Container and CUDA execution of a historical scalar replay only |
| AI Cloud Serverless Jobs, EU North | `aijob-e00he632g06bw1kxtv`, completed | H100 CUDA audit of twelve original photographs, each photograph's own mask, and saved neutral/candidate renders | Real source-bound pixel processing, 24 verified heatmaps, per-view report; diagnostic, not asset promotion |

The B300 smoke log showed `NVIDIA B300 SXM6`, driver `580.173.02`, CUDA `13.0`, and 275040 MiB device memory. Nebius reported container execution from `2026-10-22T04:53:37.930440300Z` to `04:53:38.510811462Z` for that job. Container time does not establish total billable time or cost.

The H100 job used `gpu-h100-sxm`, preset `1gpu-16vcpu-200gb`, and reported `COMPLETED`. Its container executed from `2026-10-22T06:29:32.920721850Z` to `06:29:35.243632245Z`; the log showed `NVIDIA H100 80GB HBM3` and PyTorch `2.9.0+cu130`. The CUDA arithmetic took about 0.223 seconds. It averaged twelve scores that were already computed locally: control RGB MAE `15.548863728841145`, candidate `15.553794066111246`, candidate better in six views. The candidate remained `not_promoted`. **The job did not load the original photos or masks, reconstruct geometry, generate a material, render a camera view, or produce a new asset.** The label `ARGUS_GPU_PROOF` in its historical log is overbroad; it identifies this receipt, not proof of a cloud product stage.

The replay input SHA-256 is `f743ce1b0833c60aa1e38d67b76995cf4484d1d7062977a7e817a9c5e9cebf5e`; the local source report SHA-256 is `538724b63ea6e47de53ce27da432d91bac4a31d69920a5d8c70d133302dd0e59`. The uploaded entrypoint SHA-256 was `d18f9be69ea5562004ce5579fd7fe71ba07b6d42b64017434bc614aed5968c53`. Provider exports on Robin's Windows machine are `local review bundle/argus-nebius-h100-job.json` (SHA-256 `58c072e5ba3a61d13c793054c9a426a9b6fb99b49202d3d2aa5e2ae73838ba30`) and adjacent `argus-nebius-h100-logs.txt` (SHA-256 `e616b48aa504d8104a47c2d373c5a036295b95bdd51fb564046ca9873f62140a`). Preserve these records; do not repeat this billable scalar test. `ship_argus_nebius_gpu_job.py` is retained only to reproduce its historical entrypoint and now requires `--historical-replay`.

## First meaningful cloud asset job: completed diagnostic

The H100 audit accepted 12 source-bound bundles via a private Nebius Object Storage mount. Its manifest SHA-256 was `ff39dedacb5cd9f1dd606d7f9723fabc8c56f46bbb16d91cc535b30f9c8bb5df`. Nebius reported `COMPLETED`, with execution from `2026-10-22T19:09:44.149409247Z` to `19:09:49.994890703Z`. The output report SHA-256 was `db2521329dee3223394275c2eb4f641b9cb9aa1856135b521dfe9318947c18f5`; 24 PNG heatmaps and the JSON report were downloaded and hash verified. Its mean neutral and candidate RGB MAE values were 15.41145 and 15.40383 respectively at 360 × 542, with eight of twelve views lower-error for the candidate. The change is very small and the sampling contract differs from the full-resolution local gate, so the material remains diagnostic. The cost method and caveats are in the linked report.

## Next cloud asset stage

The next Nebius AI Cloud stage can move asset generation or full-resolution camera parity to the cloud. `ship_gof_backend.py` is an actual CUDA surface-evaluation path, but its pinned extension and scene dependencies must be packaged and qualified on the selected cloud GPU before it is a runnable job. The fox material and original-camera comparison in `ship_fox_material_model.py` and `ship_fox_material_camera_compare.py` is the output-parity target: it loads the GLB, texture atlases, original photographs and masks, twelve camera transforms, and light receipt, then writes images and a per-view report. Those two scripts use NumPy/SciPy and trimesh, with no demonstrated CUDA path; merely moving them into a GPU container would not prove GPU processing. Their current runner uses a pinned local Docker image and Ahab-mounted working tree. Neither runtime and source package is ready for Nebius yet.

For a future paid asset-generation job, package one immutable input manifest covering its exact scene data, photos, masks, cameras, script revision, pinned extension, and container image. Preserve source ownership and camera bindings. Write a retrievable mesh, field, or rendered images plus a report; logs alone are insufficient. Compare that artifact against a frozen local result using the same cameras and record metrics and differences. Include the fox fixed-camera material comparison if the cloud stage produces a compatible surface or asset; do not substitute its CPU execution for CUDA evidence.

Do not promote geometry or materials merely because the cloud job succeeds. Preserve the local report's `diagnostic` status and the existing failed or conditional qualification gates. The cloud acceptance gate is reproducible execution and artifact parity, not a claim that the fox is already game ready. The exact AI Cloud billable cost remains unknown until a billing record is captured. Token Factory balance and access do not establish AI Cloud credits. Published-rate arithmetic and measured usage are in the linked report.

Official setup: [Nebius Serverless AI Jobs quick start](https://docs.nebius.com/serverless/quickstart/jobs) and [CLI setup](https://docs.nebius.com/cli/configure).