Tensor Earthquake Classifier
Outcome-trained earthquake classifier on the Earth State Tensor. The worldwide version was retired from the ledger on 2026-10-10 after it lost to the base rate on its own forward record. The per-region version stops writing on 2026-10-25; a July 2026 test measured it at chance.
What it does
Takes the Earth State Tensor as input, projects onto a 50-dim eigenmode + non-zero-feature subspace, and predicts whether a global M>=5.5 earthquake will occur within the next 3 days. Trained on real USGS outcome labels for each historical tensor snapshot, not on synthetic distributions.
Physics basis
The tensor fuses Schumann resonance, telluric currents, GNSS strain, IGETS superconducting gravimeters, cosmic-ray muon flux and many other signals. The classifier learned which combinations came before quakes from the real outcome history. Tested on later data it had not seen, it did no better than chance. (The seismic-velocity-change input, dv/v, is paused since 2026-10-14 while we revalidate it.)
When it fires
The per-region version logs to the ledger as tensor_eq_per_region until 2026-10-25. The worldwide version (tensor_eq_global) has logged nothing since 2026-10-03.
What the customer receives
- Retired: no new deliveries
- Its past ledger rows stay on the public record, unedited
Operational numbers (live)
Engagement paths
Per-query, subscription, territorial-exclusive, and royalty-on-find structures are all available. Specific commercial terms are scoped after a technical-fit conversation.
Honest caveats
- 67 snapshots with real outcomes is small; temporal correlation between neighboring snapshots means effective independent sample count is lower.
- Per-region classifiers need 6-12 months of tensor data before some sparse regions become trainable; no accuracy figure is quoted.
- Confidence gate is set loosely (>=0.40) because tensor probabilities are calibrated via isotonic CalibratedClassifierCV -- not the raw logistic output.
Michael Jessop: michael@deepmapai.com · partner portal