ST-000: High Learning-Rate Sanity
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.
Mechanismβ
The instability is induced by violating the local stability margin:
or, when the true Hessian is available:
Expected Behaviorβ
| Variant | Expected result |
|---|---|
| float, no throttle | diverges or oscillates |
| float, global throttle | stabilizes by reducing |
| fixed-point, no throttle | diverges faster or hits rails |
| fixed-point, global throttle | stabilizes if the quantization interval is feasible |
Main Plotsβ
- loss vs step,
- vs step,
- vs step,
- vs step,
- 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.