Apple and Google swap the model inside your phone with each OS update and ship no changelog. Underfoot freezes one 28-prompt suite, runs it five times per build and diffs the answers. All 140 macOS 26.5.1 runs came back byte-identical, which is the finding that makes drift measurable at all; the iOS 26.3.1 simulator failed every generation and flipped a guardrail refusal. Once a build is superseded it can never be measured again.