Teacher models fail 57% of trials on the τ 2-bench, wasting expensive data during tool-calling distillation. Apple introduced PROOF-Gen to recover these failures by optimizing trajectories instead of discarding them. This method turns near-misses into training signals.