Twelve LLM batch jobs in twelve containers, all ephemeral

You want to A/B/C/.../L test twelve prompt variants against the same evaluation set of 5,000 examples. Each variant takes about three hours. Running them sequentially is a day and a half. Running them in twelve threads on one machine fights for tokens-per-minute on the same API key.

Twelve containers, one POST each to /api/v1/daemon/quick-start with ttl: 14400 (four hours). Each ephemeral program runs one variant against the eval set, writes results to a shared hoody-files path, exits. After four hours each ephemeral config self-deletes. Twelve idle containers go back to costing nothing.

You GET twelve status URLs the next morning. All STOPPED, exited cleanly. Results are in the shared file. No leftover config, no ssh sessions to clean up, no zombie processes from a screen detach.

Twelve POSTs, twelve workers, twelve self-cleaning containers: the experiment costs nothing once it's over.