Generic early-warning signals are at chance under a matched false-alarm protocol

Brain, heart, grid, climate and molecular data — one operational bar

Most published early-warning evidence rests on a retrospective per-record trend test. This paper asks the operational question instead: at a fixed false-alarm budget, does a detector fire on transitions more often than on no-transition controls? Across scalp-EEG seizures, cardiac AF onsets, power-grid growing oscillations, palaeoclimate transitions and dynamical-network-biomarker indices, no generic detector reaches significance — including the canonical Dakos AR(1) baseline on its own data.

Preprint of record: doi:10.5281/zenodo.22113062 (Zenodo, CC-BY 4.0, 26 August 2026). Under review at Journal of the Royal Society Interface. Copy on Academia.edu.

Hosted PDF is byte-identical to the Zenodo record (MD5 28f7c9000563cd3788c2d4ce06f06ff2).

A domain-specific modal-growth detector and an eigenvalue regime map

The grid positive result, its deployment bounds, and when the form transfers

The companion positive result: a detector reading the canonical wide-area-monitoring quantity — the exponential growth rate of the most unstable bus's voltage-deviation envelope — clears the identical matched-false-alarm bar on PSML generator trips where every generic detector stays at chance. Streaming recalibration, ISO-NE and WECC 240-bus cross-dataset transfer, and external validation against small-signal eigenvalues bound the claim and yield a regime map: envelope growth on oscillatory modes, autocorrelation on non-oscillatory ones.

Preprint of record: doi:10.5281/zenodo.22113116 (Zenodo, CC-BY 4.0, 26 August 2026). Under review at IEEE Transactions on Power Systems. Copy on Academia.edu.

Hosted PDF is byte-identical to the Zenodo record (MD5 6027788713f79ee72b8079d8a4190605).

Reproducibility

Evidence-bound manuscripts

LaTeX sources, build instructions, citation metadata and submission records live under papers/submissions/ in the public repository. The sealed evidence artefacts behind every number are committed under examples/real_data/; integrity tests recompute each content hash from the payload alone.