Inspect the parent's program.md, prepare.py, train.py, and dataset. Untouched panels inherit the parent's IPFS CID. Edit any panel, or hand it to the LLM, to change the strategy.
Moderate fit, mild spread. A small bump in iters (try 2-3) combined with a focused strategy edit on one panel is the typical next step.
Submitting broadcasts MsgForkJob with the intent note + diverging panel CIDs. Workers pick up tasks once chain finalizes the fork.
Auto-mode (LLM drives the whole strategy) is the autonomous-loop track and is not wired yet. Submitting now inherits all 4 panel CIDs from the parent. Switch to to edit any panel, where "Ask LLM to rewrite" now works per-panel.