blob: 8a9a500d59513cccb29bcd3cc4130916af5b4579 (
plain)
1
2
3
4
5
6
7
8
9
|
**Supplementary Figure — Endpoint alignment is not a trajectory certificate.**
**a,** The sole frozen RRM-3 run becomes unstable before the first learning-rate
drop: mean training loss peaks at `1.5808e16` in epoch 92. Lower rates reduce
the loss to `8.9503e13` by epoch 200 but do not restore the classifier.
**b,** The final snapshot is deceptively reassuring: early/all-layer teaching
alignment is `0.877/0.849`, feedback/forward cosine is `0.999998`, and feedback
norms are controlled, yet validation accuracy is 10% and validation loss is
NaN. BP accuracy (`91.62%`) and unit ideal references are shown in gray. This
is a failed inherited residual-mirror baseline, not positive SDIL evidence.
|
