INTERACTIVE FIGURE 13The rotating frame — why layer-k probes die at layer k±4, and the one-map fix
depth (block) 0
Schematic of the measured law, with the measured constants. The feature's information is present at every block (its length barely changes — R² stays ≥0.9 for a probe refit at each depth), but its direction turns ~45° per block, so a probe frozen where it was fit reads a vanishing projection: naive transfer collapses within 2–3 blocks and goes negative — worse than guessing the mean, because the probe now reads other features' coordinates. The alignment map is fit blind to the feature (one linear map per layer pair, unsupervised), yet transporting the probe through it holds the read at 96–99% of the refit ceiling across the entire stack — because all features ride one coherently drifting frame, and one map re-expresses the whole frame. Measured anchors: adjacent rotation 45.5° ± 8°, end-to-end cosine 0.04 (T70 material) and 0.055 (Qwen bracket depth), naive transfer R² ≈ −1, aligned +0.93/+0.96.