Live demo — Modal Head-to-Head¶
Open the interactive demo — the certified comparison behind the grid modal-growth detector, charted from the sealed record.
What the demo shows¶
On 90 real generator-trip instabilities of the PSML 23-bus corpus (Zheng et al. 2021), the domain-specific modal envelope-growth detector leads 36 transitions at a matched 10% false alarm (label-permutation p = 0.0001, and 24/45 on the held-out half, p = 0.0002) — while every generic early-warning detector, including the classical critical-slowing-down baseline, stays statistically at chance on the identical split. The chart shows each detector's led count against its own expected-by-chance mark, with the p-values sealed alongside.
The comparison is fair by construction: identical segments and calibration,
disturbance-type labels independent of the scored statistic, disclosed
data-quality drops, and a pre-registered operating point validated on a
held-out half. The whole record is one content-addressed artefact
(examples/real_data/psml_modal_growth/) guarded by an integrity test.
Claim boundary¶
The certificate is corpus-specific: the numeric operating point does not port across datasets (measured on ISO-NE captures — see the Grid Sentinel Replay demo), so deployment calibrates per system on its own ambient data. "Generic detectors at chance" is a sealed statement about this task and operating point, not a blanket dismissal of those methods elsewhere.