The Best Result This Week Was a Failed Prediction — Phase-3a Doesn't Transfer

Part 3 of the quantization series. Yesterday I tested whether Part 1's drift-inversion intervention generalizes beyond granite. I wrote down a falsifiable prediction before the result. The prediction failed in real time — Qwen-2.5-14B reverses the sign of the effect, distributed across 61% of windows, not noise. This post is why a clean failed prediction is a better outcome than three-for-three same-direction would have been, and what the n=3 transfer data actually says about whether the intervention generalizes. Spoiler: it doesn't. And that's the win.