The study of which longitudinal observations, sensing schedules, representations, and governance constraints are sufficient to support reliable target-specific prediction. The name proposes a research program, not an established field: it draws on personal sensing, lifelogging, ubiquitous and mobile sensing, egocentric vision, activity recognition, behavioural forecasting, HCI, wearable computing, privacy-preserving machine learning, and information and decision theory. The contribution is the measurement link — TargetSpace connects an observation choice to the gated, target-specific predictive skill it produces, so sensing questions get answered by experiment rather than assumption.
Version 1.0 — pre-pilot protocol proposal. Synthetic harness only. No human-subject results.
The limiting factor for personal AI is increasingly not only model capability but observation quality. A stronger model cannot indefinitely infer target-specific state from evidence that does not contain it, and richer evidence is not automatically useful: quality, continuity, attribution, coverage, and timing constrain what any model can learn about a particular person. Personal-AI evaluation is incomplete when model performance is measured without characterizing the evidence available to the model. TargetSpace treats observation quality as a first-class experimental variable.
The least burdensome governed evidence configuration that preserves a pre-specified level of gated, target-specific predictive skill. It is task-dependent — for one outcome calendar metadata may suffice, for another audio may be necessary, for another sparse images add real lift, and for some outcomes no acceptable configuration exists at all, which is itself a benchmark result. When less invasive sensing produces equivalent gated skill, the smaller configuration wins the comparison: the objective is minimum sufficient governed observation, not maximal surveillance.
Evidence efficiency is a family of experimental ratios, not a score: EE(Eₖ | Eₐ) = ΔSkillR2 ÷ Cost(Eₖ) — the incremental gated skill an added modality, sensor, or schedule contributes, per disclosed unit of cost (joules, worn hours, valid capture hours, captured bytes, or manual activations). Only gated skill counts: sealed, calibrated, permutation-passing, coverage-disclosed, architecture-matched. Privacy is deliberately a vector, never a single score. And the optimal system is not the one with the highest raw skill — TargetSpace does not assume that richer sensing is better; it asks which evidence configurations remain non-dominated after predictive lift and observation cost are measured together.
Non-wear, charging gaps, and consent pauses are structured, not random. Explicit missingness logs keep “no event” distinguishable from “no observation” and protect calibration analysis from silent imputation.
Devices change the behaviour they measure — visible cameras, indicators, prompts, charging rituals. Forgettability is an experimental property: awareness reports, interaction counts, removal frequency, early-vs-late behaviour drift. Inconspicuousness never overrides consent or recording law.
Hypothesis, not finding: a small fraction of observed time may carry a disproportionate share of transition-relevant information. Routine and transition strata are reported separately; any capture trigger must be frozen before evaluation — adaptive sensing that peeks at outcomes is leakage, not efficiency.
Fix the task set. Vary the sensing arm — tier, modality, schedule, placement, device. Seal every forecast before its outcome exists. Score against the R1 population-prior and R2 own-routine baselines, gate on calibration and wrong-target permutation, and attribute any lift to the evidence that produced it via the evidence-tier ablation. The result is a measured answer to a sensing question — including the null answers, which are just as informative.