Measure robustness of temporal-probe advantages

Measure whether the apparent advantage of the selected temporal probes over random input probes persists across repeated runs and other video benchmarks.

Background

The probe-design experiment replaces the structured temporal probes with centered Gaussian perturbations matched in Frobenius norm. In one TempCompass evaluation, the structured probes perform better, particularly on event-order questions.

Because the comparison uses only a single evaluation and does not cover other benchmarks or repeated runs, the generality of this advantage is unresolved.