score-studio

Score Studio

Use the scorestudio_* tools to operate Score Studio. Start with scorestudio_whoami when the organization is unknown.

Operating rules

1. Inspect before mutating. Resolve the organization, dataset version, model version, evaluation run, workflow, or deployment rather than guessing an ID.
2. Explain cost or external effects before starting training, evaluation, or a workflow run. Preserve returned durable identifiers.
3. Treat dataset and model versions as immutable provenance. Never silently substitute a newer version.
4. Compare evaluation runs only when task, dataset version, class mapping, metric configuration, and evaluation purpose are compatible.
5. Release claims require measured evaluation evidence. A successful training task or artifact export is not proof of improvement.
6. A workflow response may contain deferred blocks. Report its execution mode and deferred work instead of describing preview as production completion.
7. Verify a deployment and retain the returned health evidence before calling it production-ready.
8. Do not expose access tokens, private media, provider credentials, signed URLs, or raw secrets in chat or logs.

Useful sequences

• Audit readiness: list datasets → inspect models → list evaluations → fetch the chosen report → list and verify deployments.
• Train: inspect the dataset/version → explain compute choice → start training → return the durable model/version identifier.
• Evaluate: resolve exact model and dataset version IDs → start evaluation → later fetch the immutable report.
• Workflow: list workflows → confirm the input object key and side effects → run → distinguish executed from deferred blocks.