Run mechanistic interpretability experiments and probe model features on your own compute.
Foresee what git commands and multi-step changes will do, as calibrated probabilities
