ST-001: Distribution / Operating-Point Shift
Purpose
This is the main deployment-realistic stress family. The model is stable under a nominal condition, then the deployed input distribution changes and the same learning rate becomes unsafe.
Mechanism
Examples:
The design rule is:
The learning rate is valid before deployment shift and becomes unsafe only after the operating condition changes.
Variants
| Variant | Purpose |
|---|---|
| no input cap, no throttle | expose the raw failure |
| input cap, no throttle | test whether capping alone is enough |
| no input cap, throttle | isolate learning-loop stabilization |
| input cap, throttle | test practical combined protection |
Controller Hypothesis
The global throttle should respond to the increased closed-loop sensitivity:
Main Plots
- loss and output error,
- input scale/drift schedule,
- ,
- and ,
- rail pressure if fixed-point is enabled,
- stability margin .
Input capping critique
Input clipping can protect the forward signal, but it does not necessarily stabilize the closed-loop weight update. Valid experiments should compare cap-only, throttle-only, and cap-plus-throttle variants.