@emollick Agree on the human baseline. The finance and law near-zeros only show up under all-or-nothing grading, while the leaderboard itself uses partial credit. A human baseline under both rules would tell us whether 0% means "can't do it" or "almost done, one step wrong".
@emollick This also explains why researchers feel it most. We build multi-month workflows (prompts, pipelines, reproducibility notes) on top of tools that were never meant to be stable. Keeping your own copy of the prompts and settings is the boring fix.