My auto research template features

• always in worktree
• git hooks reminding and protecting files
• explicit human,meta,agent files, explicit frozen files
• everything is a file, with a pipeline from idea -> worktrees -> research reports
• stronger and explicit ideation and interpreations (weak points of current agents)

My specific ML things:

• uses pueue to queue GPU jobs from multiple agents
• use jaxtyping and beartyping to document dimensions and do typechecking / shapechecking in the smoke test
• tyro/dataclass for config
• TODO ml_debug skill to (hopefully) uplift agent dignosis

See also

• https://github.com/karpathy/autoresearch/blob/master/program.md
• https://github.com/NousResearch/autonovel/blob/master/program.md

TODO key plot or log from each experiment