RESEARCH QUESTION

Can a causal evidence gate preserve ETAS-relative information gain while limiting negative transfer from a learned spatial correction?

This is not a public earthquake warning product. It is a prospective shadow test in which daily spatial rate forecasts are published before their target window and scored only after events occur.

Four forecast lines

ETAS: Frozen scientific benchmark
Safe incumbent: Renewal-frailty plus supported-neighbor residual
Fixed spatial expert: Moves mass only within the highest-ETAS support
Evidence-gated forecast: Exactly equals safe until the BF20 hurdle is exceeded

Causal gate

Completed-day fixed-versus-safe event log ratios accumulate in shadow mode. The next forecast weight is fixed as:

w = max(0, tanh((S − log 20) / 2))

The prospective state starts at S=0; retrospective evidence is not carried forward. BF20 is an operational evidence hurdle, not an unconditional 5% false-activation guarantee.

Scoring

IGPE(M, ETAS) = Σ log[λ_M(i)/λ_ETAS(i)] / N

Safety reporting also gives fixed and gated IGPE relative to the safe incumbent. Every layer preserves the same daily expected count, so the paired compensator difference is zero.

Prospective integrity

Forecast artifacts are timestamped and SHA-256 locked before the target day.
Backfill is prohibited.
The first score is provisional and the seven-day catalog snapshot is final.
The 14-day operational dry run does not count as scientific evidence.
A 365-day claim-bearing test requires a separate frozen protocol and commit.

Limitations

The method was developed after California, New Zealand, Chile, and Japan-C outcomes were opened. Those results are not independent confirmation. The retrospective gate removed large negative transfer but retained a very small Chile loss; the prospective test targets that unresolved question.