Success Criteria
ELFQuake should be judged in stages. Current outputs are engineering artifacts, not earthquake predictions.
Stage 1: End-to-End Scaffold
Success means the pipeline is reproducible and inspectable.
- A single command emits a weekly
>M2event-list CSV and JSON report. - Each event row has time, latitude, longitude, magnitude proxy, probability proxy, source scores, and warning fields.
- The output can be rendered on the Italy map without manual conversion.
- Reports record source paths, class counts, model type, calibration context, and caveats.
- Missing modalities are represented by masks, quality fields, or explicit source metadata.
Current status: mostly satisfied by trial-weekly-event-forecast.sh, learned-weekly-event-forecast.sh, and trial-forecast-map.sh.
Stage 2: Synthetic Model Utility
Success means synthetic training is better than trivial synthetic baselines.
- Learned scorer balanced accuracy exceeds
0.60on unseen synthetic episodes; within-episode temporal splits are supporting diagnostics only. - Positive and negative recall are both at least
0.40; a model that predicts almost everything positive does not pass. - At least 80% of repeated unseen-episode folds meet both recall floors, so a favorable mean cannot hide unstable episodes or initializations.
- Synthetic event-list targets have both classes, location targets, and a positive rate between
0.10and0.90. - Event-list adapters report occurrence, count, magnitude, centroid, timing, rate, and spatial-spread metrics, not only a binary score.
- Temporal train/test positive-rate drift is small enough for a meaningful chronological check; target positive-rate delta should be below
0.25before promoting synthetic temporal scores. - Forecast count calibration is within
25%of the synthetic held-out weekly event rate. - Ablations show whether direct avalanche, piezo/VLF-like, and combined features each add or remove value.
Current status: partly satisfied only as an engineering check. The original synthetic h6 table still fails with temporal positive-rate delta 0.609182. The warmed aggressive profile with WARMUP_STEPS=3000 now passes the drift gate on a nine-episode run with delta 0.187025 and 396 labeled rows, but temporal model utility remains weak. A nine-episode WARMUP_STEPS=1000 run failed the same gate with delta 0.294146.
Latest controlled within-episode check: stable name-derived initialization makes matched model variants directly comparable. Random-init piezo/VLF-only averages 0.619033 balanced accuracy over three seeds and keeps both recalls above 0.40 in every seed. Synthetic pretraining lowers it to 0.534272; current self-supervision therefore does not add downstream value on this synthetic target.
Unseen-episode check: leave-one-episode-out evaluation over nine episodes and three model seeds averages 0.578712, ranges from 0.275641 to 0.758730, and only 14 of 27 folds keep both recalls above 0.40. Stage 2 therefore fails despite the stronger within-episode mean.
Fixed-seed ensemble check: averaging the three predeclared model seeds raises unseen-episode mean balanced accuracy to 0.632634, but only 6 of 9 episodes meet both recall floors. A fixed 0.5 threshold scores 0.601361 and passes the same 6 episodes, so threshold calibration is not the main instability.
Alignment checks: shortening the target to h3 scores 0.580991; expanding h6 context from 12 to 60 simulation minutes scores 0.512103; retaining simple spatial sensor aggregates scores 0.544096. None passes the gate, and the original h6/12-minute mean-only representation remains the best tested configuration.
Late-fusion check: the controlled random-init anchored full and direct-only variants average 0.609241 and 0.606385, below the matched 0.619033 piezo/VLF anchor. Disabling the direct branch raises its mean to 0.611281, so direct avalanche has not demonstrated incremental value.
Stage 3: Real Supervised Readiness
Success means real data can support supervised evaluation.
- Real VLF-aligned rows contain both positive and negative labels in at least one scope.
- Training rows occur strictly before validation rows.
- A seismic-only historical-rate baseline is computed for the same target and period.
- VLF and astronomical features have documented coverage and timestamp alignment.
Current status: blocked. Latest real aligned rows are one-class: all-Italy 149/0, central Italy 0/149, plus 129 future rows pending in each scope. Labels now require the source event catalog to cover the complete target horizon, so these counts no longer include catalog-gap rows.
Stage 4: Predictive Claim Threshold
Success means the system demonstrates value on held-out real data.
- The multimodal model beats seismic-only and historical-rate baselines on held-out time periods.
- Improvements survive ablations: VLF and/or astronomical features must add value when included and lose value when removed.
- Probability calibration is acceptable by bucket, not only by rank metrics.
- Results are reproducible from recorded source pulls and command scripts.
Do not claim earthquake prediction capability before Stage 4 is satisfied.