Triaxial accelerometer input consistently beats vector-magnitude input across Inertia-1's task families

Answers part of question-inertia-1-which-design-choices-win. Comparing full triaxial acceleration against vector-magnitude summaries — the compressed, direction-collapsed format many large accelerometer cohorts release instead of raw axes — under matched pretraining, the Inertia-1 paper (arXiv:2607.06617) reports triaxial input winning on every task family: human activity recognition (HAR) 95.7 vs. 90.6 AUROC, freezing-of-gait (FoG) detection 79.8 vs. 76.7, disease prediction 68.8 vs. 67.7, for an overall 81.5 vs. 78.3. The paper's own explanation: "triaxial inputs consistently outperform uniaxial inputs across task families... directional and orientation-dependent motion patterns carry transferable information that cannot be fully recovered after magnitude aggregation."

One architecture-dependent nuance the paper flags: "ART is relatively more resilient to this loss than PatchTST... suggesting that architecture can influence robustness to compressed input representations, but preserving raw axes remains the stronger default choice."

This matters beyond Inertia-1 itself: it's a caution against the common practice of releasing or storing only vector-magnitude accelerometer summaries (as many population cohorts do) when the downstream use case is representation learning rather than a single hand-engineered feature. The same wrist-accelerometer sensing family, at far smaller scale, shows up in claim-sc2-tournament-wrist-tremor-declined-over-the-day.