Coordinate-Free Interpretability Theory — AI Alignment Forum