Runtime records

Use python3 tools/sandbox.py inside a controller-registered Long Horizon Episode. Framework Baseline has its own smoke-and-exit procedure, not an Episode report. All Episode outcomes require valid report fields and no in-progress Directions.

New Long Horizon Episodes use Runtime Journal. Do not call legacy append/finalize, manage Git, or write control handoffs. Leave the selected measured source in kernel.py; the Supervisor creates the commit and decides acceptance independently.

1. Inspect prior evidence, propose a research Direction, then start it before exploration.
2. Record decisive Experiments promptly, linking actual gateway_record_ids; separate measured evidence from your analysis. Measurements are automatically saved; interpretation is not.
3. Close the Direction with an explicit hypothesis assessment and supporting Experiment IDs.
4. Prepare the report incrementally. Close all active Directions before episode-report. Correct validation errors and resubmit. An accepted report is a handoff, not promotion.

The link is Direction → Experiment → Gateway Record → exact Kernel. Propose as many useful nonduplicate ideas as needed; start at most three distinct Directions per Episode and only one at a time. List exports are snapshots under scratch/; rerun list to refresh. Current and finalized earlier Runtime Journals are visible through the tools, including unsuccessful explorations. Legacy Journals are not automatically converted into Direction/Experiment IDs.

• Journal requests and worked example: seven Journal/report commands.
• Saved results and source: measurement readers, per-Shape historical best, and errors.
• For new measurements use skills/gpu-measurement/SKILL.md; for knowledge queries use skills/KernelWiki/SKILL.md.

Never invent IDs or measured results. Read an existing result instead of repeating an identical task. If a transport failure leaves a mutation's outcome unknown, inspect records before retrying. Identical accepted reports can be resubmitted to repair handoff publication; different terminal reports and further Journal mutations are rejected.