The thesis
A disengaged human produces labels that look exactly like engaged ones. A human with a frontier model open in the next tab produces labels that are, by construction, indistinguishable from human judgment. Output review cannot catch this — the output is the disguise. Quality is not a property of the label. It is a property of the human, at the moment of judgment.
The signal
An involuntary motion biomarker — micro-dynamics of a brief phone gesture that a person does not consciously control. Scored server-side against a per-user calibrated baseline. A language model cannot produce it. A checked-out human cannot fake into it. Engagement isn't asked about — it's measured.
Reproducibility
Sensie publishes an open eval harness so partners can reproduce the core claims on their own data: github.com/sensie-app/sensie-eval-harness. Scope: signal extraction reference, calibration protocol, and reproduction of the 83.6% post-calibration accuracy figure.
Evidence
9 PhD-led research trials · 18,000+ sessions · 83.6% post-calibration accuracy · 2 granted US patents + 1 filing.