Skip to main content

ST-000: High Learning-Rate Sanity

Status:Preliminary
Tags:Ablation TestSanityGlobal Throttle
Last modified:

Purpose

This is the minimal controllability sanity check. We intentionally choose a learning rate that is too large and verify that the global throttle can reduce the effective step size.

Mechanism​

The instability is induced by violating the local stability margin:

Ξ·Ct>2.\eta C_t > 2.

or, when the true Hessian is available:

ηλmax⁑(Ht)>2.\eta\lambda_{\max}(H_t)>2.

Expected Behavior​

VariantExpected result
float, no throttlediverges or oscillates
float, global throttlestabilizes by reducing Ξ·teff=Ξ±tΞ·\eta_t^{\mathrm{eff}}=\alpha_t\eta
fixed-point, no throttlediverges faster or hits rails
fixed-point, global throttlestabilizes if the quantization interval is feasible

Main Plots​

  • loss vs step,
  • Ξ±t\alpha_t vs step,
  • Ξ·teff\eta_t^{\mathrm{eff}} vs step,
  • CtctrlC_t^{\mathrm{ctrl}} vs step,
  • Ξ±tΞ·Ctctrl\alpha_t\eta C_t^{\mathrm{ctrl}} vs step.
Interpretation

This test proves that the controller can reduce an unsafe learning rate. It does not prove the deployment claim by itself. It should be treated as a sanity check.