Agent Blame
A pipeline of AI steps gave a wrong answer. Swap any step for a better version, see whether the answer comes out right, and see how much of the blame each step gets.
Loading the replays…
The browser does not run the pipelines. Every combination of fixes was replayed beforehand by the Python package, with the randomness held fixed; the page looks those runs up and computes the blame itself, with the same formulas.
How it's built, results and limits: How it works →