Experiments
The Snapshot Bug and 1D Collapse Retraction
The Snapshot Bug and 1D Collapse Retraction
Retraction. V20, V25, and V26 evolution loops reset hidden states to zero BEFORE saving snapshots. All hidden state analysis was on zero vectors. The "1D energy counter" finding was completely wrong.
Corrected V22-V24 data: effective rank 5.1-7.3 (not 1-3), energy = -3.6 to -4.6 (not 1.0), PC1 variance 25-38% (not 95-100%). Hidden states are moderately rich and do NOT encode energy linearly.
Lesson: When hidden states appear degenerate, first verify you're analyzing actual post-cycle states, not zeroed reset buffers.