ST-007: Legacy κ Projection Mismatch
Purpose
This is not a natural external failure mode. It is a comparison against the original 🚂ENABOL stabilizer.
Mechanism
The old row/column mechanisms constrain representational gain and rail pressure, but the applied update may become:
In general:
The global throttle instead applies:
so the intended update direction is preserved before fixed-point quantization effects.
Questions
- Does row/column projection keep rails safe?
- Does it rotate or distort the learning trajectory?
- Is the loss stable but learning degraded?
- Does global throttle recover stability with less geometry distortion?
Main Plots
- loss,
- rail pressure,
- row/column scale events,
- update cosine / phase,
- per-layer update norm,
- global throttle comparison.
Priority
This belongs late in the ablation ladder. It explains why the project moved away from aggressive row/column projection, but it should not be presented as the main way standard fixed-point training fails.