Skip to content

History

Revisions

  • models-employed: add the real neural experiment record (5.5) — strain context + calibration + gate win The seismicity-only gate ~0; the GNSS-strain context + the rate_cal calibration fix -> the neural beats ETAS on the in-loop gate (+0.075 nats, N-test passes, forecast 218.9 vs 248 observed). With the honesty that defines the product: single-window, NOT pseudo-prospective; under validation; never published on one window.

    @fsantibanezleal fsantibanezleal committed Jun 19, 2026
    6a125f8
  • evaluation: add the real results section (back-analysis + multi-model benchmark, live) Real leakage-free numbers from the harness, rendered live on the Back-analysis page: 195-tile ETAS on 97k events; global field vs null +0.0835 nats/eq, high-low bias gap +0.0184; the multi-model benchmark (ETAS best, naive ensemble underperforms ETAS) with the honest finding + links to the experiment register and the cited evidence base. Negative/zero results kept.

    @fsantibanezleal fsantibanezleal committed Jun 18, 2026
    f140aa5
  • changelog: headline results are now LIVE — update status table + central numbers The status-at-a-glance was stale (it said full training / real back-analysis were 'pending'). They are done and live: 97k-event global fit (195 ETAS tiles), the 8-view pseudo-prospective back-analysis, global context contribution 0.0835 nats/eq vs null, high-low bias gap +0.0184, neural trained (gate ~0 until covariates), efficiency 13x. Calibration (N-test) flagged as the honest open item.

    @fsantibanezleal fsantibanezleal committed Jun 18, 2026
    f0f8569
  • wiki: restructure into per-topic sub-pages (deep) + Data section (sources + features) 19 per-topic sub-pages (one deep page per classical model: Gutenberg-Richter, Omori-Utsu, ETAS, Reasenberg-Jones, STEP, EEPAS, smoothed seismicity, BPT/renewal, rate-and-state+Coulomb; one per ML method: temporal point processes, RMTPP, Neural Hawkes, Transformer Hawkes, RECAST/FERN, CNN spatial (DeVries+Mignan), GNN/LSTM, detection vs forecasting) + Data-Sources (the sources) + Data-Types-and-Features (the data types + engineered features) with mermaid/equations. Models-Classical/ML become indexes; References consolidated; Sidebar/Home rebuilt as the tree. Per-topic = depth (Felipe's directive).

    @fsantibanezleal fsantibanezleal committed Jun 17, 2026
    27fb987
  • wiki: initial technical record + persistent disclaimer footer 13 pages mapping the deep research: Home, _Sidebar, Glossary, Methodology-History, Honest-Limits, Models-Classical, Models-ML, Models-Employed, Data-Sources, Technical-Architecture, Pipeline, Evaluation-and-Tests, Changelog-and-Progress. _Footer.md puts the forecast-not-prediction / not-for-life-safety disclaimer at the foot of every page. Public-safe (leak-scanned). Next: restructure into per-topic sub-pages for depth + add the Data (sources + features) section.

    @fsantibanezleal fsantibanezleal committed Jun 17, 2026
    ce195d5
  • Initial Home page

    @fsantibanezleal fsantibanezleal committed Jun 17, 2026
    0fcaeb5