Valkyrie distinguishes evaluated results from the files produced by an agent.
Full syntax is in the run results, run outputs, and run output reference, and in client.runs.results for Python.
Download results
Results hold the scores and evaluation. Without --path, the CLI writes ./results-<run-id>.json.
Pass a task subset to recompute final_score over just those tasks. The stored full result is unchanged:
Combining --s3 with a task subset overwrites the canonical S3 result key with that subset view. Run the command again without filters to restore the full result.
Download outputs
Outputs are the files the agent produced. Fetch all of them, or only the tasks you care about:
To pull one file or directory instead of the whole run, name a path inside the run’s benchmark prefix:
The output directory defaults to ./<run-id>.