rl-standard-job-cleanup

name	rl-standard-job-cleanup
description	Preserve, validate, and hand off a terminal standard MarinSkyRL run backed by dataset rows and programmatic rewards, without an agent harness or per-trial artifacts. Use after completion or failure, or before an authorized cancellation. Use rl-agentic-job-cleanup for agent trajectories.

Clean up a standard RL run

This workflow preserves evidence. It does not authorize cancellation, publication, registration, deletion, or database mutation.

Workflow

1. Classify the run. Read the resolved configuration, not the name. Confirm that the run has no agent harness, sandbox, or per-trial artifacts.
2. Establish terminal state. Pair controller state with durable outputs and the run's configured step budget. Preserve the first causal error for failed jobs.
3. Capture the full chain. Collect logs from every restart or resume plus training metrics, checkpoints, export metadata, and the resolved launch command/configuration. Do not treat absent trials or verifier artifacts as missing evidence for a standard run.
4. Validate candidates. Check shard completeness, step consistency, configuration/tokenizer metadata, metric coverage, and the last saved step. Derive model size from exported tensors.
5. Select only by declared policy. If the experiment defines no selection criterion, report all valid candidates and leave selection unresolved. State the metric window and chain coverage for any selected artifact.
6. Hand off or publish only with explicit authority. Verify destination, visibility, provenance, and a clean secret scan. Confirm whether registration applies before creating a row.
7. Reclaim only with separate authority and only after the destination has been verified.

Before an authorized cancellation, capture live process state, per-rank device utilization, and in-flight outputs that will disappear. Do not terminate an unexplained reproducing failure before recording the evidence needed to diagnose it.

Completion record

Report terminal state, source revision, resolved configuration, preserved locations, validation results, selected artifact when authorized, first failure cause, resume safety, and remaining work.