Work

Scale Invariant Covariant Kinematic Auditing for Environmental Tipping Points: A 30 Site Year Flash Drought Benchmark and Falsification of Direct Solar Surface Coupling

Year: 2026
Type: preprint
Abstract: Most early warning systems look in the rear-view mirror, waiting for 14 to 30 day trailing averages to drift far enough to trigger an alarm. That habit creates a structural Diagnostic Lag: the warning arrives after the regime shift has already taken hold. We test whether CovariantSentinelAuditor, a kinematic detector originally built for non-equilibrium information geometry and validated on NASA flight crew physiology, can break that lag zero-shot. Retaining its exact invariant structural weights and phase-gating boundaries , we evaluate it untouched across two real-world geophysical problems. First, we test direct solar-to-surface atmospheric coupling (HELIOS_01 ATMOS_02) across 3 years (2022–2024) of hourly NASA SPDF OMNI2 solar-wind telemetry and Copernicus ERA5 London data (17 persistent high-pressure blocks). A 50-shift circular permutation test falsifies direct solar-surface coupling , demonstrating why solar-terrestrial attribution requires strict null controls. However, when we strip the solar term away, the uncoupled atmospheric auditor (uncoupled_atmos) catches 100.0% of blocks (17/17) versus 76.5% for a legacy 14-day trailing Z-score at essentially the same alert rate (16.37% vs 16.72%). Second, we apply the identical core to agro-hydrological flash droughts (BIO_05) across 30 site-years (2015–2024; 38 onsets) in East Anglia (UK), Iowa (US), and Castile-La Mancha (Spain). Using only surface soil moisture (0–7 cm), reference evapotranspiration (ET0), and vapour pressure deficit (VPD) - blind to the root-zone soil moisture (7–28 cm) defining ground-truth labels and the legacy baseline - bio05_thinning_or_veto achieves 0.789 recall [95% CI: 0.667, 0.896], 0.260 F1 [0.190, 0.327], and 16.00 days mean lead [13.52, 18.21]. Under 1,000-replicate site-year block bootstrapping, this delivers a statistically significant +3.44-day lead-time gain [95% CI: +0.15, +6.55] over a legacy 30-day root-zone Z-score (12.56 days, 0.711 recall, 0.241 F1). Pairing the kinematic cue with a causal physical drying gate (bio05_twostage_confirm, bio05_drying_thinning_or_veto) exposes a clean operational Pareto frontier, cutting alert load to 8.12% while nearly doubling early-warning yield per false-alert day (1.87 vs 0.96). Without re-tuning, a detector built for AI latent manifolds and validated on flight-crew physiology spots geophysical regime shifts earlier—not by averaging longer, but by watching how the trajectory itself bends. Reproducible Code & Artefacts: https://github.com/Regsiren/covariant-sentinel-earth
Author: Olaiya, Remi
Language: English
Open Access status: green
OpenAlex ID: W7237200192
Cites: 2
Cited by:
Related to: 10