Live demo — The Honest Test

Open the interactive demo — the four-domain falsification of generic early warning, charted from the committed record.

What the demo shows

The best SCPN suite member and the classical Dakos et al. 2008 AR(1)-Kendall-tau detector were evaluated on real transitions in four domains (EEG seizures, cardiac AF onsets, grid instabilities, palaeoclimate tippings) at a matched 10% false alarm with label-permutation significance. No detector reached significance in any domain; the strongest signal anywhere is the classical detector on EEG at p = 0.067. The chart shows each detector's led count per domain against its expected-by-chance mark.

Publishing a negative result about our own detectors is deliberate: the same harness that demoted the generic suite is the one that certified the modal detector which does clear the bar (see the Modal Head-to-Head demo), and it is the harness offered for certifying any detector on any data. A certification you can trust is one that is allowed to say no.

Claim boundary

Corpora are small (6-12 transitions per domain) — the honest statement is "no detector demonstrated significance here", not "early warning is impossible". One classical competitor was implemented; the EEG corpus is single-subject. The committed record is examples/real_data/head_to_head_ar1_kendall/; the publication-form study is Early-Warning: Honest False-Alarm Evaluation.