Watches, bands, rings, and pendants already collect the one of the longest continuous physiological and mobility records of any product category — heart rate, sleep, motion, location, and in some devices ambient audio. Yet that record is mostly rendered as descriptive scores: readiness, strain, sleep quality. The open measurement question is whether any of it yields calibrated, target-specific forecasts of what the wearer does next — and TargetSpace is a protocol for testing exactly that.
Version 1.0 — pre-pilot protocol proposal. Synthetic harness only. No human-subject results.
Personal-AI products are converging along one ladder: recording → structured memory → longitudinal evidence → target-specific predictive capability → validated assistance, with the predictive-capability rung measured by TargetSpace. Wearables entered the ladder from the opposite end to most software: they began with continuous longitudinal sensing and only later added interpretation. The record is deep, but a record is not a model — retrieval and summary scores describe what the wearer did; they do not, by themselves, forecast what the wearer will do.
Current rung: longitudinal evidence. Wearables hold multi-year physiological and mobility histories per wearer, surfaced as trends and scores. Next rung: target-specific predictive capability — demonstrated not by richer dashboards but by sealed forecasts of the wearer's next observable transitions that beat the wearer's own routine and collapse under target permutation.
Wearables instantiate the upper evidence tiers of the flagship TS-Personal track. Whether each tier adds forecasting skill is a hypothesis the protocol tests, never an assumption.
GPS and visit traces, movement patterns, commute regularity. The candidate signal for transitions that show up first in where the wearer is.
Heart rate, heart-rate variability, sleep stages, motion and activity intensity. The tier only wearables can supply continuously.
Pendant- and watch-class devices with microphones add ambient audio. Its marginal value over L5–L6 is measured by evidence-tier ablation, not presumed.
Passive sensing is the working hypothesis of this whole category — that observation beats asking. TargetSpace treats it as exactly that: a hypothesis, tested tier by tier against the self-report channel, which remains a biased but legitimate auxiliary baseline.
At a sealed decision point, forecast whether a recurring health-adjacent commitment — a scheduled workout, a planned sleep window — resolves as {complete, defer, cancel, replace}. The forecast is timestamped and sealed before the outcome exists; resolution is deterministic from the device's own record (the session occurred, the sleep window was kept, within pre-registered tolerances).
Scored with proper scoring rules against the R1 population-prior baseline and the R2 own-routine baseline; calibration is checked over the full distribution; and the permutation specificity gate requires skill to collapse when forecasts are scored against a different wearer. An evidence-tier ablation (physiology off, mobility off, audio off) reports what each stream actually buys.
Pick one recurring commitment type. Run your forecaster A-B over the same sealed tasks with a single feature toggled — physiology on vs. off, or mobility on vs. off — and compare skill. One toggle, one answer.
Raw sensor data never leaves the company. Only sealed forecasts, resolved outcomes, and aggregate metrics are exported for scoring. The protocol audits skill, not data.
One leaderboard row: skill vs. R1, skill vs. R2, calibration, permutation result — reported per evidence tier, so the marginal value of each sensor stream is a number, not a claim.
Do physiological signals add target-specific forecasting lift beyond routine, or do they largely re-encode the same regularities the R2 own-routine baseline already captures?
Which mobility features — location entropy, visit patterns, commute regularity — carry forward-looking information about upcoming transitions, rather than merely describing past ones?
What battery and wear-time envelope maximizes longitudinal coverage, given that charging gaps and non-wear periods may correlate with exactly the events being forecast?
Can low-power sensors (accelerometer, optical heart rate) substitute for high-power ones for most transition types, and for which task families does the ablation show the difference matters?
How should sensor dropout and non-wear periods be represented in the evidence manifest so that missingness is scored honestly rather than silently imputed away?
None of these questions has an empirical answer yet. TargetSpace v1.0 is a pre-pilot protocol proposal with a synthetic harness; these are the questions a first wearables pilot is designed to make answerable.