Skip to main content

ST-007: Legacy κ Projection Mismatch

Status:Preliminary
Tags:Ablation TestKappa BudgetingLegacy
Last modified:

Purpose

This is not a natural external failure mode. It is a comparison against the original 🚂ENABOL stabilizer.

Mechanism​

The old row/column κ\kappa mechanisms constrain representational gain and rail pressure, but the applied update may become:

Δθtactual=Πκ(θt−ηGt)−θt.\Delta\theta_t^{\mathrm{actual}} = \Pi_\kappa \left( \theta_t-\eta G_t \right) -\theta_t.

In general:

Δθtactual∦−ηGt.\Delta\theta_t^{\mathrm{actual}} \not\parallel -\eta G_t.

The global throttle instead applies:

Δθtctrl=αtΔθtraw,\Delta\theta_t^{\mathrm{ctrl}} = \alpha_t \Delta\theta_t^{\mathrm{raw}},

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.