Research portal

Withdrawn 2026-10-14. The reconstructed archive this recipe loads (retro tensor v1) was later found to leak up to about a day of future data into each hourly snapshot: a value was taken from the daily record nearest in time, including the day still to come. That can produce exactly a "flagged in the hours before" result, so this replay is not evidence of an advance signal. The builder was fixed (causal from v5 on); whether a real signal exists is a research question we will re-test there. The steps below still reproduce the original number.

1. Load the real reconstructed 2005–2018 Earth-state tensor (data/retro_tensor.db, 122,712 hourly snapshots built from public archives: NOAA F10.7, NMDB cosmic ray, USGS FDSN seismicity, WDC Kyoto Dst, IGETS gravity, INTERMAGNET).
2. Load tensor_eq_retro_v2 (temporal) — trained on only the first 80% of the timeline (≤ 2016-03-14). All six test events are in the held-out last 20%, so this is genuinely out-of-sample.
3. Run scripts/backtest_six_event_oos_replay.py — rebuilds the training feature matrix, scores every snapshot → P(any M6.5+ globally within 24h), and reads each event's probability + percentile at T−3h … T−72h.
4. "RED" = probability ≥ p90 (top decile of all 122,712 snapshots).

Expected result: 6-for-6 RED by T−3h — Kaikōura M7.8, Bougainville M7.9, Tehuantepec M8.2, Gulf of Alaska M7.9, Fiji M8.2, Palu M7.5 — each at the 98.1–99.7th percentile at T−3h…T−12h.

Honest caveats: the lead is hours, not days — the signal fires strongest at T−3h…T−12h and is driven substantially by the elevated-seismicity / immediate-foreshock state (real ETAS-style clustering), not exotic multi-physics precursors. This is a global M6.5+/24h nowcast, not per-event location pinpointing. Model skill = AUC 0.806 out-of-sample temporal CV.

See reports/research_notes/six_event_oos_replay_2026-07-06.json (sha256 0ab94eef2c21) for per-event lead curves.

1. Clone cross_physics/convergence_gate_predictor.py.
2. Run the 30-day baseline bootstrap: python -m cross_physics.convergence_gate_predictor --cycle --no-log × 10.
3. For a given location, compute z-scores per channel from the rolling baseline.
4. Verify: if per-channel FPR ≤ 30%, triple-gated FPR ≤ 2.7% by multiplicative compounding.

Note: a low triple-gated false-positive rate follows by construction only if the channels are independent. Measured precision is not yet established, so no precision figure is quoted.

1. Access data/earth_state_tensor.db (tensor snapshots since Apr 14, 2026).
2. For each snapshot epoch, query USGS for M≥5.5 globally in next 3 days — binary label.
3. Project tensor onto nonzero-active-feature subset (~298 dims).
4. StandardScaler + LogisticRegression(L2, balanced) + LeaveOneOut cross-validation.
5. ROC AUC across LOO predictions.

Withdrawn 2026-10-14. These steps give a high cross-validated score on overlapping snapshots that does not survive a temporal holdout (the per-region family measured at chance, HYP-9fb3d2fa); the model was retired from the ledger 2026-10-10. We no longer quote the number. Script: scripts/train_tensor_eq_classifier.py.