Registered gates for "One ECG, three routes to the breath": results and characterisation of two failures

Both gates registered in 10.17605/OSF.IO/K4ZAX were read first, in the registered order, and both failed. This report discharges the reporting obligation that registration creates. GATE 1 (beat detection against Kubios HRV Scientific, ten recordings, 600 s each): six of ten studies reached the registered 99% of our beats matched within ±50 ms; the criterion is nine. The range is 52.83–99.86% and bimodal, with nothing between 97.8% and 99.2%. GATE 2 / R5 (order-shuffled RR control): registered to span zero at every window length; at 15 s it gives +0.0124 [+0.0014, +0.0257]. A corrected re-run failed harder. The cause is a conceptual error in the control itself: shuffling the order of RR intervals preserves the times of the beats, and at short windows those times are the respiratory signal, so the control preserves what it exists to destroy. Characterisation. The disagreement is episodic rather than systematic, and no systematic timing offset among mutually detected beats was observed — which excludes a constant shift but not a shared error both detectors make. It is also not visible from our own RR series. A second, genuinely independent implementation (plain NeuroKit2) then answered the question a two-way comparison cannot: in three of the four failing recordings our detector is the outlier, so GATE 1's failure is an engineering defect rather than a general property of clinical ECG. Erratum (v1.1, 12 August 2026). Version 1.0 described that check as a third detector with "no shared code with either party". That was false for one party: the study's own pipeline is NeuroKit2 (ecg_clean + ecg_peaks) plus correct_artifacts=True, so the check involved two algorithm families, not three independent detectors. The verdicts and the conclusion are unchanged; the warrant is corrected: Kubios vs plain NeuroKit is the independent corroboration, and ours vs plain NeuroKit is a controlled single-variable experiment that localises the divergence to the artifact-correction argument. An exploratory retest with the argument off reaches 7/10 at the registered criterion and does not reverse the registered 6/10. Full record: deviations register D22 (updated file in this version). No registered hypothesis (R1–R6) is interpreted. The deviations log D1–D21 is included, and records the defects found in the checking apparatus as well as in the analysis.