k=1 Is A Methodological Trap

Boundary-capability models can reverse conclusions between single-shot and repeated measurements.

Current Evidence

In the archived compression cross-check, a Claude Haiku condition changed direction when rerun at k=3. The practical rule for this dataset is simple: k=1 is acceptable only when a model is obviously far from the boundary. For interesting effects, use k >= 3.

Consequence

New contributions require k >= 3.