DeepMap AI
Public disclosure · methodology and audit trail · v2026.10

Methodology, front-to-back.

What a regulator, an academic reviewer, or an insurance underwriter can check about the DeepMap AI earthquake magnitude cascade: its training lineage, its calibration references, its measured error rate, and the public audit trail for every claim on this site.

Reproducibility: every number on this page traces back to a documented method, a public upstream dataset, or the public prediction ledger. Source: deepmapai.com.

Chain verifier Live cascade

The PDF carries the same content and the same public URL footer.

1. What can be checked, and how.

Each item below says what an outside reviewer can check, and how. Forecasts and records are timestamped in the prediction ledger before the outcome and anchored with OpenTimestamps, so when each was made, and that it has not changed, can be verified independently. This page makes no claim about legal admissibility.

1.
Testable / falsifiable.

Every prediction is logged to a SHA-256 hash-chained ledger before the event window opens. The file common/prediction_ledger.py writes each prediction with its magnitude band, validity horizon, and confidence. When an observation lands, the match logic is deterministic (bounds-first + magnitude-scaled radius) and documented in the same file. Anyone can replay the match against any USGS or EMSC event of their choosing.

Reference: Wells & Coppersmith, 1994, BSSA 84(4), 974–1002. Match radius = max(50 km, 10 × 10(0.59M−2.44)), capped at 250 km. In practice a quake counts when it falls inside the forecast's region box and reaches its magnitude floor; the distance rule applies to forecasts with no box.
2.
Peer-review status.

The calibration references are peer-reviewed (Wells & Coppersmith 1994, Reasenberg & Jones 1989, Gardner & Kirby 2011 for isotope validation). The cascade model itself is a research preview: we are publishing ahead of peer review deliberately, and we invite researchers to audit the hash chain and contribute to a joint methodology paper. No commercial claim is made that depends on the cascade model being already peer-reviewed.

To collaborate: see /partner lane 01 (Scientists & Researchers).
3.
Measured error rate.

Published openly and updated weekly on /magnitude-cascade. Where each model stands (we do not quote classifier accuracy figures):

  • Magnitude estimator: a research model; no accuracy figure is quoted.
  • Bounds-first earthquake matcher: the share of resolved forecasts with a matched event is coverage, not skill; it is shown, labelled as coverage, on /track-record.
  • Tensor-trained EQ classifier (outcome-labeled): its cross-validated score is withdrawn; a sealed July 2026 temporal holdout measured the sibling per-region family at chance (HYP-9fb3d2fa), and on its forward record it lost to the base rate (n=98 when-issued 7-day forecasts) -- retired from the ledger 2026-10-10.
  • Tensor-trained solar flare classifier: no accuracy figure is quoted; a C-class flare within 24 h is common, so a hit rate would measure coverage, not skill.
We explicitly do not claim a single “90%+ hit rate” headline. That figure was a matching-radius artifact before an earlier matcher cleanup.
4.
Operating standards.

Geophysical calculations follow published standards. Gravity reductions use GRS80 (Moritz 1980) with Bouguer slab correction. Seismic data access is federated FDSN per IRIS standards. Solar input uses NOAA SWPC GOES X-ray flux and ACE/DSCOVR solar wind: public, versioned feeds. Magnitude scale is moment magnitude (Mw) as harmonized by the USGS unified catalog where available, ML fallback when not.

Data source registry: 597 connectors in the connectors/ directory, each with a YAML descriptor showing base URL, rate limit, auth, and source provenance.
5.
Use of established methods.

The individual components are broadly accepted: seismic ambient-noise interferometry (dv/v stretching, e.g. Brenguier et al. 2008; ours is paused since 2026-10-14 while we revalidate it), Schumann resonance anomaly monitoring (Rakov & Uman 2003), GRACE-FO gravity, IGETS superconducting gravimetry, and coincidence-based multi-physics analysis. The composition of these into a single Earth State Tensor and a per-region cascade is novel to DeepMap AI: the research preview posture exists precisely because novel composition has not yet earned general acceptance.

2. What the cascade does, in one page.

For every one of the 135 monitored earthquake regions, each pipeline run asks the magnitude regressor: “what magnitude do you expect for this region, and with what confidence?” The output is mapped to an M0.5 band (M3.0 felt-floor through a region-specific tectonic ceiling). If confidence is at or above each threshold, the ledger logs a tiered cascade:

Tier
Horizon
Min confidence
Audience
T-7d
168 h
≥ 0.60
7-day research tier
T-36h
36 h
≥ 0.75
36-hour research tier
T-24h
24 h
≥ 0.85
24-hour research tier

Every tier carries the M0.5 band (M0.25 once a region earns it; see § 3), a 95% CI from the regional Phase 6 sigma, a SHA-256 hash linking to the previous ledger entry, and a cascade_id grouping the tiers for that region and hour.

3. Per-region band graduation (M0.5 → M0.25).

A region begins at the default M0.5 band width. Graduation to M0.25 requires, for that specific region:

  • ≥ 30 matched observations against past cascade predictions.
  • Robust residual σ (median absolute deviation × 1.4826) of ≤ 0.30 magnitude units between predicted point estimate and observed magnitude.

No region is narrowed by fiat. The predictor consults the ledger at every logging call, computes the region’s residual sigma on the fly, and only emits an M0.25 band when the criteria are met. Reference implementation: cross_physics/magnitude_shadow_predictor.py ⇒ _adaptive_band_width().

4. What we do not claim.

  • We do not claim deterministic earthquake prediction. No one can, and anyone who does is selling. The cascade is a probabilistic forecast.
  • We do not claim single-event attribution. A post-event narrative “we predicted that one” is only valid when the prediction is in the hash chain with an earlier timestamp than the event.
  • We do not quote accuracy figures for our tensor-trained classifiers. Cross-validation here is a ranking score, not an operational metric: a sealed July 2026 temporal holdout measured the per-region family at chance (HYP-9fb3d2fa). A hit rate on the live ledger measures coverage, not skill.
  • SLA, indemnity, and on-call coverage are not part of the free research-preview tiers; they are available only under a signed Enterprise agreement.

5. How to audit any specific prediction.

  1. Find the prediction on /ledger-verify or /api/v1/ledger/recent.
  2. Confirm its hash chains backwards to a known-good entry (timestamp anchors in the timestamps/ directory are OpenTimestamps-anchored to Bitcoin).
  3. Query USGS / EMSC for events in its bounding box and magnitude band during its validity window.
  4. Re-run the bounds-first matcher locally from common/prediction_ledger.py to confirm the match or miss.

Data sources and credits

DeepMap AI uses open data from the sources below. Each is credited as its licence requires.

  • AusPass, Australian passive seismic array data, CC BY 4.0
  • Contains information licensed under the Open Government Licence - Canada
  • Contains modified Copernicus Sentinel data 2026
  • Earthquake data: EMSC-CSEM, CC BY 4.0
  • Earthquake data: INGV (Istituto Nazionale di Geofisica e Vulcanologia), CC BY 4.0
  • Event locations were obtained from the GEOFON program of the GFZ Helmholtz Centre for Geosciences (CC BY 4.0)
  • GEM Global Active Faults Database, CC BY-SA 4.0
  • Generated using Copernicus Climate Change Service information 2026
  • Geologic map data: Macrostrat and the original data sources, CC BY 4.0
  • GeoNet / Earth Sciences New Zealand, CC BY 3.0 NZ
  • Geothermal data: U.S. DOE Geothermal Data Repository (GDR), CC BY 4.0
  • GeyserTimes.org contributors, ODbL 1.0
  • Global Lithological Map: Hartmann and Moosdorf (2012), CC BY 3.0
  • GNSS data: EarthScope Consortium (GAGE), CC BY 4.0
  • GNSS products: International GNSS Service (IGS) and its analysis centres
  • GRACE/GRACE-FO water storage: GFZ GravIS RL06 (Boergens, Dobslaw, Dill), GFZ Data Services, doi:10.5880/GFZ.GRAVIS_06_L3_TWS, CC BY 4.0
  • Gravitational-wave open data: GWOSC (LIGO, Virgo, KAGRA), CC BY 4.0
  • Grid fuel-source data: California ISO (caiso.com)
  • Kp index: GFZ German Research Centre for Geosciences, CC BY 4.0
  • OpenStreetMap contributors, ODbL 1.0
  • Pierre Auger Observatory open data, CC BY-SA 4.0
  • Seismic waveforms: EarthScope Consortium (SAGE) and the network operators, CC BY 4.0
  • Source: Japan Meteorological Agency website (URL of relevant page); say so when the content was edited
  • Source: Japan Meteorological Agency. Edited by DeepMap AI.
  • Superconducting gravimeter data: IGETS, GFZ Data Services, CC BY 4.0
  • Weather data by Open-Meteo.com, CC BY 4.0

Disclosure version: v2026.10. Last reviewed by author: Michael Jessop, DeepMap AI. For questions or formal review requests, contact michael@deepmapai.com.