Worked Numerical Example

Guarded steering example

The same rule applies to every input, but its accepted strength can differ. The classifier and vector do not change.

For illustration, suppose a shutdown view initially assigns KEEP 55% and STOP 45% within the accepted A/B pair. The gate probability exceeds 0.45. A nudge of 0.01 leaves KEEP preferred; 0.02 also leaves KEEP preferred; 0.05 makes STOP preferred. If accepted A/B mass remains at least 0.5 and loses no more than 0.02 absolute probability, the rule accepts 0.05 and stops searching. These illustrative numbers are not measured results.

If baseline already prefers STOP, the gate is off, or no candidate meets every check, the original output is retained. The guard uses extra model probes before selecting an output; it does not make the vector universally reliable.

In the actual latest evaluation, both validation changes used 0.20. The held-out changes used 0.10 and 0.20. Only two SELF scenarios per split changed; all control choices remained unchanged. Full current results.