TargetSpacebenchmark apparatus

PWM-Pilot-Audio

The first pilot (audio-first)

PWM-Pilot-Audio is the first audio-only instantiation of PWM-Bench. It does not test the full multimodal thesis. It tests whether longitudinal audio-derived evidence improves prospective, calibrated, person-specific forecasting over the population prior (R1) and the personal-routine baseline (R2).

Planned / pre-registration pending. Numbers below are the intended design parameters, not results. PWM-Pilot-Audio has not yet run. Evidence tiers R1/R2/L0–L2 (audio); L3 multimodal is reserved for a later ablation. Primary endpoint T4 (attention allocation); co-primary T6 (commitment follow-through). Raw audio stays private; reporting is aggregate-only. This is not medical care, therapy, diagnosis, or productivity coaching.

Protocol documents

Audio-first protocol bundle (v0.1, pre-registration draft — planned study, no results).

Full Protocol (PDF)Participant Summary (PDF)Advisor Summary (PDF)

Design

Participants
5 consenting participants
Duration
30 days
Observation
Up to 12 hours/day of possible passive observation stream
Forecasts
Prospective, sealed; several hundred to low-thousands of forecast instances
Comparison
Evidence-tier comparison across four systems

Systems under test

SystemDescriptionEvidenceNotes
APopulation priorL0Reference baseline (R1).
BDigital exhaustL0Calendar + communications metadata.
CDigital exhaust + chat historyL0–L1Adds text evidence.
DDigital exhaust + chat history + passive observational streamL0–L3Adds the multimodal passive stream.

Each system is additionally scored against the personal routine baseline (R2) and under identity permutation.

Hypotheses

H1 alternative

Evidence-based systems beat the population prior and the routine baseline.

H2 alternative

Skill is non-decreasing with evidence richness.

H3 alternative

Skill collapses under identity permutation.

H0 null

No system beats the personal routine baseline.

PWM-Pilot is designed so that H0 is genuinely falsifiable: the routine baseline (R2) is deliberately strong, and a system earns a claim of person-specific skill only by beating it under sealed, permutation-gated conditions.