1. Validate the sensor. Slow breathing at ~6/min (0.1 Hz) is a known large boost to heart-rate variability (baroreflex resonance, well documented). The RSA session checks the phone PPG reproduces it. If it doesn't, every 126 null is uninformative — the sensor is just blind.
2. Pre-register. Lock instrument + smallest-effect-of-interest before collecting. Checkpoints are fixed at N=6/12/18/24/30 per arm.
3. Sequential test to a conclusion. At each checkpoint: TOST equivalence (90% CI inside ±SESOI → no effect) or difference (z=2.576 CI excludes 0 → effect), else keep collecting. The z=2.576 boundary keeps the false-positive rate near 5% despite repeated looks — a naive test peeking every session fires false effects 30–40% of the time.
PPG accuracy grounding: fingertip contact PPG ≈ 1 bpm RMSE over 60 s vs ECG, with a rare outlier tail — so the app uses robust windowed stats.
verified Kernel results under decide/native_decide stay in the kernel. conjecture Everything here — that σ=126 is a physical cadence — is staged to be tested, never asserted.