Experiment Launch

Launch rules

• run_manifest.json is launch-only and immutable after start.
• Use schema_version: 1 in the manifest and initial events.
• Preflight approval and long-run approval are separate checks. A passed preflight makes the run eligible to launch; it does not replace explicit long-run approval.
• Do not start a long run unless the approved preflight evidence matches the intended code, command, environment class, and device class closely enough to be relevant.
• Do not start a run without explicit metric parser rules, expected outputs, and checkpoint discovery rules.
• Do not auto-resume an older run from this skill. Use experiment-resume.
• If the intended run root already contains active lease or status state, stop and resolve the collision instead of overwriting it.
• Shell wrappers must preserve the child exit code in both the wrapper exit status and exit_code.txt.
• Do not handwrite a zsh wrapper with status=$?; status is read-only in zsh.
• Sequential batch launchers must not raw-read monitor-owned status.json without retry handling.
• Sequential batch launchers must write batch_status.json atomically and preserve batch.stderr.log on internal failure.

Output

After launch, provide:

• experiment ID
• run root
• PID
• initial status
• the preflight evidence that made the launch eligible
• the next required handoff

References and templates

• Use assets/run-manifest-template.json.
• Use references/runtime-contract.md.
• Use references/launch-checklist.md.
• Use scripts/write_run_wrapper.py.
• Use scripts/smoke_test_run_wrapper.py.
• Use scripts/launch_unlearning_lmc_bridge_batch.py for sequential bridge batches.
• Use scripts/smoke_test_unlearning_lmc_bridge_batch.py after changing the tracked batch launcher.