Skip to main content
Run an Evolve task quality check to review task files against a rubric. Submit an archive or a published dataset, then poll the returned check until its task checks settle. Examples reuse connection setup. This request checks one task from a published dataset:
Success returns 202 Check, including its id and one results entry per selected task. Set CHECK_ID to that id. This starts billable checker work; it does not run the evaluation agent on the task.

Routes

Result fields are in analysis and check results; language equivalents and full signatures are in check methods.

Request fields

Supply exactly one of archive and dataset. For an archive, replace the dataset field with -F 'archive=@tasks.tar.gz'. All configuration fields are optional: Read GET /api/checks/defaults, optionally with agent, for the current agent, model, rubric, prompt, reasoning effort, and provider. See checker configuration for template details. Omitting concurrency uses the organization ceiling; task checks and trace analyses share that ceiling.

Read and download

The check status is queued, running, or completed. A completed check means its task checks settled; inspect each result’s status, checks, and failure before treating it as successful. A checker that produces an invalid result is retried at most once; a budget timeout is not retried. List filters are optional scope (my, shared, org), repeatable status (queued, running, completed), dataset, limit, cursor, sortField=created, and sortDirection=asc|desc. Default scope is my; default order is newest first. See pagination.
A check id downloads the whole check report and task-check folders. A task-check id from results[].id downloads only that task-check folder. The run must be settled. For sandbox files and logs, use the shared filesystem routes with resource prefix /api/checks/{checkId}/tasks/{taskCheckId}.

Share or delete

Share requests use link: true, emails: [...], or both. They follow the job sharing rules, including creator-only management and a maximum of 50 email grants. New email shares send invitations.
Send the same body to /unshare to revoke the link. GET /shares reads its current state. To permanently remove a check you created:
This deletes its task checks and stored results. A dataset used as the source remains available.

Errors

Empty selections return 400 no_checkable_tasks. Invalid archives or configuration return invalid_archive, invalid_input, or invalid_rubric. Upload capacity can return 413 upload_too_large or 429 too_many_concurrent_check_uploads; the latter has no fixed retry delay. Download and deletion refuse active checks with 409 check_not_terminal; the same code can mean results are still being finalized. Deletion and share management require the creator. See check error codes.
Last modified on November 9, 2026