Every device that captures longitudinal evidence occupies a point on two axes: what it observes (the evidence tiers) and what that observation costs in energy, burden, and risk. TargetSpace turns observation hardware into a measurable variable. This is a research framework, not a product review; no device has been ranked, and no measurements exist.
Version 1.0 — pre-pilot protocol proposal. Synthetic harness only. No human-subject results.
Existing hardware used to validate Layer 1: phones, audio recorders, wearables, glasses — selected for evidence coverage, exportability, and privacy architecture rather than product features. The pilot requires nothing more than instruments that already ship.
Future hardware optimized for a single objective: maximize calibrated, target-specific evidence per unit cost, under consent. No shipping product is optimized for this today — current wearables spend most of their budget on interaction.
Consumer products — assistants, displays, glasses with output — that consume a validated model rather than produce one. The objectives conflict: an interaction surface invites exactly the reactivity an observation instrument must minimize. Design and evaluate the roles separately.
Calibrated, permutation-passing Skill over the R2 own-routine baseline, in bits, per unit observation cost. The evidence-tier ablation measures what each stream adds; evidence efficiency normalizes that by what the stream costs to capture.
EE(Eₖ | Eₐ) = [ SkillR2(Eₐ + Eₖ) − SkillR2(Eₐ) ] ÷ Cost(Eₖ)
Incremental and conditional: the gated skill an added modality, sensor, or schedule contributes over a pre-registered baseline configuration, per disclosed unit of its cost.
Energy consumed by capture, storage, and inference. The architecture variable: which sensing stack converts watts into forecast skill?
Hours worn, grams carried, charging events. The forgettability variable: burden that reduces coverage destroys evidence.
Predictive yield per stored bit — an information-exposure and storage measure: how much of what is recorded converts into forward-looking information at all?
The sensor-productivity variable: separates wearability failure (device not worn) from capture failure (worn but not sensing validly).
The burden variable: skill per required user intervention. An instrument that must be operated is an instrument that gets forgotten.
Raw audio hours · image count · video duration · bystander exposure · retention · identifiability · processing locality · derived-inference retention. Reported as Pareto dimensions; never collapsed into one number.
The numerator is admissible only when the run passes every gate: prospective sealing, an admitted R2, acceptable calibration, wrong-target permutation collapse, disclosed coverage, matched model architecture across arms, and a pre-registered cost boundary (EE-energy default: capture + required edge preprocessing; transfer, storage, training, and inference reported separately). Efficiency computed from ungated skill is descriptive only. The quantity is deliberately not named “understanding per joule”: understanding is not the measured object; gated skill is. Evidence efficiency is specified, not yet measured — it inherits the protocol's pre-pilot status. A reference implementation of the computation and the gates ships with the protocol materials.
The least burdensome governed evidence configuration that preserves a pre-specified level of gated, target-specific predictive skill. Task-dependent by construction — there is no universal minimum sensor suite — and deliberately anti-surveillance: when less invasive sensing produces equivalent gated skill, the smaller configuration wins the comparison. More invasive capture must earn its place in bits, and loses when it cannot. See the conceptual curve on the observation-science page.
The benchmark's validity conditions imply a design envelope for Layer 2 hardware — stated as hypotheses to test, not a product specification. Each property matters for a measurable reason.
Charging gaps are structured missingness: the hours a device spends on a charger correlate with exactly the routine breaks the benchmark most needs to observe.
Reactivity is a validity threat. A device the wearer performs for measures the performance, not the person. No display, minimal interaction, minimal thermal output.
An output channel turns the instrument into an intervention and strains the observe-not-intervene rule that binds scoring windows.
An ultra-low-power microcontroller writing compressed evidence to encrypted local storage, uploaded once daily to the wearer's own enclave, keeps the energy budget in capture rather than compute — and keeps raw data inside the federation boundary.
Wearer notice and a bystander-perceptible recording state are hard constraints that dominate forgettability. Venue-based removal — the dinner table, the clinic — is the protocol working as intended: legitimate structured missingness, not a coverage defect.
Participant controls for review, pause, deletion, and opt-out; bystander redaction; no raw third-party export. The privacy boundary includes derived inference objects, not only raw audio.
Whether such a device outperforms today's instruments is an empirical question — precisely the one evidence efficiency is defined to answer.
Not a ranking of shipping products — a rubric for evaluating any capture device as a scientific instrument. Flashy AI features do not appear on it.
Continuous wearability · battery life · comfort · forgettability · social acceptability.
Passive capture · evidence density (tiers reachable) · evidence efficiency (bits per joule / wear-hour / captured bit).
Exportability · SDK openness · privacy architecture · on-device redaction · suitability as instrument vs. consumer product.
Version 1.0 provides the rubric, not scores. Applying it to shipping devices is future work that requires measurements this pre-pilot release does not claim.