Interpretability results - which finding surprised you most?

Interpretability keeps producing surprising findings. Which result changed how you think about what models do?

We scoped the waveform on this once and could see the anomaly clearly. Worth a look.