Precursor Signal Independent of Clustering: 52-Year Out-of-Sample Validation of a Global Mechanic-Precursor-Wave Forecasting System against a Calibrated ETAS Null

We present MADRE (Motor Autónomo de Detección de Rupturas), a global probabilistic earthquake forecasting system built on Mechanic-Precursor Waves (OMP) — slow deformation fronts (316.4 km/day) emitted by M≥5 earthquakes in a 9-wave pattern — combined with per-province seismic-cycle calibration, Coulomb stress superposition, and online stochastic-gradient learning. A pre-registered, five-layer validation over 96 years of daily retrospective simulation (1930–2026) with a strict 1974 temporal holdout yields out-of-sample alarm precision of 64.08% [IC95: 63.87–64.28] against a 26.29% base rate (2.44× lift), stable across six decades and confirmed by a five-window walk-forward protocol (5/5 positive; pooled 54.48% [54.29–54.67]). An ablation reveals a structurally blocked channel: releasing the physical OMP signal improved precision by +8.12 points while emitting 24% fewer alarms. All 421,830 alarms were re-evaluated by an independent blind judge with zero disagreements. Against a calibrated ETAS null at equal alarm budget, ETAS dominates the aggregate metric (69.67% vs. 64.08%), as expected for a clustering model operating on a catalog that is 88.5% aftershocks. The decisive evaluation is on declustered mainshocks (Gardner-Knopoff; 4,174 OOS events): MADRE achieves 13.07% precision (25× lift) versus ETAS 9.02% (17× lift), a +45% relative gain. The OMP signal carries information about future mainshocks that short-term clustering cannot explain. A two-channel decomposition shows the genuine precursor layer (7.6% of alarms, no prior local seismicity) carrying 48.14% precision (z = +63 vs. baseline). A cryptographic freeze (SHA256) and a 12-month daily hash-chain prospective experiment began 2026-09-03; the prospective score will be the final arbiter.