← Về thư mục
📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / workflows / remote-compute / README.md

workflows/remote-compute/

submit -> poll -> harvest against a real shell — Send the heavy work out and bring the results and their evidence back. The remote script executes in a real shell.

Steps: open_session, remote_job Permissions: compute:ssh Declared artifacts: scores.csv

File Purpose
workflow.json The versioned manifest: steps, permissions, declared artifacts, failure conditions, and the cases below. Version 1.0.0. JSON rather than YAML for the same reason the core is, and versioned because a benchmark whose cases can change silently measures nothing across time.

Cases

Case Declared outcome What it pins
remote-compute/harvest success The declared output really comes back
remote-compute/unwritten-output success Promised but not produced must count as a failure

Failure conditions the manifest declares