Always-on pendants, wearable recorders, and semi-passive ambient devices capture speech, prosody, turn-taking, and timing at unusual scale for a consumer category. What the category lacks is a shared way to show that this stream buys anything beyond recall: today there is no standard test of whether continuous audio improves calibrated, target-specific forecasts of what the wearer does next.
Version 1.0 — pre-pilot protocol proposal. Synthetic harness only. No human-subject results.
The industry is climbing one ladder: recording → structured memory → longitudinal evidence → target-specific predictive capability → validated assistance, with the predictive-capability rung measured by TargetSpace. Passive audio devices clear the first three rungs by form factor alone — continuous capture makes weeks of history the default, not an achievement. The fourth rung is not supplied by hardware: a complete record of the target is not a model of the target, and memory is not understanding — a searchable transcript surfaces what was said, not what happens next.
Current rung: longitudinal evidence. The category already holds long, dense, temporally ordered evidence. Next rung: target-specific predictive capability — earned only when that evidence demonstrably improves sealed, calibrated forecasts specific to the wearer.
TS-Personal — the flagship track, currently instantiated in a synthetic harness only — orders evidence into tiers and prices each tier's marginal contribution. Passive audio enters at L2: speech content, prosody, turn-taking, and timing, layered over an L0–L1 exhaust floor of timestamps, device and app events, calendar and message metadata, with self-report retained as an auxiliary channel and baseline. That the L2 stream adds calibrated lift over the floor is the category's central claim — and in TargetSpace it is a hypothesis to be tested by evidence-tier ablation, never an assumption: the same sealed tasks are scored with and without the audio stream, and the difference is the stream's measured value.
Given the target's evidence up to seal time t, forecast the probability that a specific incoming message receives no substantive reply by resolution time r. The forecast is sealed before the outcome exists; the answer space is discrete under a pre-registered definition of “substantive”; resolution is deterministic from the message log at r.
Scoring follows the standard battery: proper scores against the R1 population-prior baseline and the R2 own-routine baseline, calibration reported alongside skill, and the permutation specificity gate — skill must collapse when forecasts are matched to the wrong target. The evidence-tier ablation then reports the same row with L2 audio removed, isolating what listening actually bought.
An A-B over the same sealed tasks: the audio feature on versus off, everything else held fixed. The difference in skill over R1 and R2 is the audio stream's value — measured, not asserted.
Raw audio never leaves the company. Only sealed forecasts, resolved outcomes, and aggregate metrics are exported — the protocol audits skill without ever touching recordings.
One leaderboard row: skill vs R1, skill vs R2, calibration, permutation result — reported per evidence tier, so the L2 audio contribution is visible on its own line.
Does calibrated lift saturate at minutes of captured audio per day, or does it continue to grow under continuous capture? The sufficiency curve decides the form factor.
Can event-triggered or scheduled sparse sampling match continuous capture on the same sealed tasks, at a fraction of the energy cost and bystander burden?
How do bystander redaction and on-device filtering change evidence value? Privacy-preserving pipelines should be scored, so their cost in bits is known rather than guessed.
Do prosody and turn-taking add lift over transcript content alone? If not, text-tier capture may suffice and the acoustic channel can be dropped.
How should gaps — device off, out of range — be encoded so that missingness neither leaks outcome information into the forecast nor biases it against low-coverage targets?