CalmSense v1.0.0

First stable release: subject-independent stress detection benchmark on WESAD (Leave-One-Subject-Out, 15 subjects) with leakage-free calibration and few-shot personalization. Includes the in-browser ONNX demo and a fully offline reproducible pipeline via make demo.

CalmSense v0.1.0

Honest, subject-independent wearable stress detection on WESAD.

Headline (Leave-One-Subject-Out, 15 subjects): Random Forest 0.913 accuracy / 0.898 F1 (binary); the four feature models are a statistical tie (Friedman p=0.81).

Four layers of optimism quantified: subject leakage, motion confound, dataset shift, and calibration.

• Leakage-free isotonic recalibration cuts ECE 0.070 → 0.025 (Brier-gap Cohen's d = 0.91, p < 0.001).
• Few-shot personalization: ECE 0.146 → 0.069 at 20 windows.
• A wrist-only model reaches 0.893 (2 pts behind the chest strap).

Reproducibility: runs entirely in the browser via ONNX (no backend); make demo reproduces the full pipeline offline on synthetic data; result artifacts are provenance-stamped (git SHA + timestamp) and WESAD checksums are verified; CI is green.

See the README and PAPER.md for full methodology and limitations. Preliminary, N=15, no clinical claim.