bingran-you's picture
Add the task author's oracle artifacts under oracle/
176b93b verified
|
Raw History Blame Contribute Delete
1.42 kB
# Integration notes
1. Select the scene with the task's current hostname-based assignment. Pin the artifact repository revision and verify the archive SHA-256 before extraction.
2. Restore the selected saved submission. Regenerate the submission assignment metadata for the current sandbox identity. Keep the original training history as provenance.
3. Recompute reported metrics with the same metric code and numerical runtime used by the verifier. Keep predicted states and renders frozen during evaluation, and refresh artifact hashes for any metadata files that change.
4. Keep full training source under `oracle/reference/training-pipeline/`. The submitted reproduction entrypoint should run the public-only training and rollout commands, without reading hidden evaluation targets.
5. Add this dataset repository and its resolved storage routes to the task's agent denylist before exposing the artifacts through a benchmark campaign. The repository URL, API/tree/resolve-cache endpoints, mirror aliases and Xet/LFS storage routes must be covered. An Oracle download may use the trusted evaluator channel.
The published 01/24 packages are historical retained submissions. The other MOSIV scenes remain pending while main is being retrained. The MASIV recovery package does not contain a complete scene workspace. Do not treat missing scenarios as successful runs or silently narrow either task's scene pool.