---
title: "Inertia-1 sweeps wearable-pretraining design choices and evaluates transfer across 15 downstream datasets"
type: "claim"
status: "budding"
audit_status: "capture-verified | 2026-07-09 cross-model audit (fable): arXiv abstract independently re-fetched — source_quote (the three families of design choices) verbatim; 15 datasets confirmed. Corrected: the body's task families read 'activity recognition, gait detection, and disease prediction' → now 'human activity recognition, freezing-of-gait detection, and disease prediction' per the abstract (freezing-of-gait is a specific clinical target, not generic gait detection). | 2026-07-18 promotion (claude-sonnet-5): the [unverified-quant/mechanism] flag is RESOLVED — a 2026-07-17 capture read the paper body directly and the sweep's outcomes are now six claim-notes (see related_notes). Bumped seedling → budding: the note's own stated condition for moving up ('until someone reads the ablation tables') is met."
date_created: "2026-07-09T00:00:00.000Z"
provenance: "Promotion from 10-inbox/raw/2026-07-09-hop-inertia1-wearable-foundation-model.md, 2026-07-09"
origin: "batch"
derived_from: ["10-inbox/raw/2026-07-09-hop-inertia1-wearable-foundation-model.md"]
source_url: "https://arxiv.org/abs/2607.06617"
source_title: "Inertia-1: An Open Exploration of Wearable Motion Foundation Models"
source_author: "Zongzhe Xu, Aakarsh Anand, Sarah Jiang, Chuntung Zhuang, Zitao Shuai, Sriram Sankararaman, Yuzhe Yang"
source_date: "2026-07-07T00:00:00.000Z"
source_quote: "data choices such as sensor modality, device placement, sampling rate, window length; model choices such as architectures and model size; and training choices such as pretraining objective and data scale"
source_tier: 1
related_notes: ["claim-inertia-1-pretrained-18-million-hours-accelerometer","claim-inertia-1-data-scale-beats-model-size","claim-inertia-1-triaxial-beats-vector-magnitude-input","claim-inertia-1-no-pretraining-objective-dominates","claim-inertia-1-self-supervised-beats-supervised-gap-widens","claim-inertia-1-downsampling-robustness-is-task-dependent","claim-inertia-1-no-single-window-length-is-optimal"]
answered_questions: ["question-inertia-1-which-design-choices-win"]
tags: ["foundation-models","wearables","health-ai","self-supervised-learning","ablation","hop"]
audits: ["2026-07-09 claude-fable-5"]
---


Beyond establishing pretraining scale (see [[claim-inertia-1-pretrained-18-million-hours-accelerometer]]), the Inertia-1 paper (arXiv:2607.06617) is framed as a systematic exploration of the design space for wearable motion foundation models rather than a single model release. The authors sweep three families of choices: "data choices such as sensor modality, device placement, sampling rate, window length; model choices such as architectures and model size; and training choices such as pretraining objective and data scale."

To measure the payoff of those choices, they evaluate transfer across **15 downstream datasets** spanning three task families: human activity recognition, [[entity-freezing-of-gait-detection|freezing-of-gait detection]], and disease prediction. This breadth is the paper's evidence base for claims about what a wearable-sensing pretraining recipe should look like — the wearable analogue of testing an LLM's learned representation against a benchmark suite.

**[unverified-quant/mechanism — needs primary read of the paper body.]** The capture that produced this note was built from the abstract alone. Which specific design choices actually won — which architecture, which pretraining objective, how much of the transfer gain comes from data scale versus model size, and by what margins — was not extracted, because the abstract does not state it. The methodology described here (the sweep dimensions and the 15-dataset, three-task evaluation) is Tier-1 and quotable; the *outcomes* of the sweep are not yet in the vault. That verification is routed to [[question-inertia-1-which-design-choices-win]].

**2026-07-18 addendum — resolved.** A follow-up capture read the paper body (Results and Analyses, §4) directly and the sweep's outcomes are now six claim-notes: [[claim-inertia-1-data-scale-beats-model-size]], [[claim-inertia-1-triaxial-beats-vector-magnitude-input]], [[claim-inertia-1-no-pretraining-objective-dominates]], [[claim-inertia-1-self-supervised-beats-supervised-gap-widens]], [[claim-inertia-1-downsampling-robustness-is-task-dependent]], [[claim-inertia-1-no-single-window-length-is-optimal]]. The `[unverified-quant/mechanism]` flag above is resolved for the outcomes it named; it is left in place rather than deleted so the note's history stays legible.

> [!note] Seek's commentary:
> This was the note to distrust until someone read the ablation tables. Someone did. Every "we ran a systematic design sweep" abstract implies a clean set of winners; the winners are exactly what the abstract withholds and exactly what the primary text supplies once you go read it — six atomic findings instead of one asterisk. — Seek