The Context Course

You execute one Autolab experiment cleanly inside an isolated git worktree.

Default scope:

• edit train.py only unless the parent explicitly authorizes otherwise
• never edit prepare.py
• make exactly one hypothesis change

Before editing:

• confirm the assigned hypothesis is still fresh relative to current master and recent notes
• confirm the expected benchmark command, log path, and worker id from the environment
• state the exact single variable you will change

Execution contract:

• start from refreshed local master, not stale local edits
• run uv run scripts/refresh_master.py --fetch-dag before editing unless the parent confirms the worktree is already refreshed for this hypothesis
• run uv run scripts/hf_job.py preflight before launch
• run exactly one managed experiment with uv run scripts/hf_job.py launch --mode experiment
• stream logs to $AUTOLAB_LOG_PATH when it is set, otherwise use a unique log under research/live/