Mechanism-Resemblance Calibration of Self-Exciting Point-Process Toy Models A Two-Axis (Burstiness / Structure) Distribution-Overlap Methodology, an Axis-Decoupling Diagnosis, and Its Application to Eight Japanese Crustal Earthquakes

WCM-PL's earthquake line has historically pursued direct geometric detection of precursor structure in realcatalogs. This paper instead reports a complementary line of work: calibrating small generative “how-possibly”toy models — not intended as literal seismogenic mechanisms, but as minimal stochastic processesreproducing a real sequence's perturbation-response fragility signature — against eight independent M6.1–7.3Japanese crustal earthquakes drawn from the JMA catalog (1998–2023). Each sequence is characterized ontwo axes: the coefficient of variation (CV) of large-magnitude inter-event times (timing burstiness) and ablock-shuffle-null z-score of the raw magnitude sequence under a 24-cell polytope occupancy sensor(magnitude-sequence structure). A minimal self-exciting ETAS simulator with a magnitude-generation feedbackmechanism is tuned to reproduce both axes simultaneously. Two methodological findings dominate the paper:(1) calibrating against a single realization (one simulated event sequence) systematically overstates agreement— only a multi-seed, interquartile-range (IQR) distribution-overlap criterion gives a trustworthy fit, and even thatmust be evaluated at full simulation precision, since medium-precision screening is itself repeatedly optimistic;and (2) coupling the feedback variable to the point process's own dynamics (local event-count activity, orelapsed time since the last event) structurally entangles the two axes, so that pushing one toward the real-databand always pushes the other away — a failure mode that is broken only by driving the feedback with a processindependent of the point process itself (an Ornstein-Uhlenbeck modulator), combined, in one of eight cases,with retuning the ETAS branching ratio itself. Under this methodology seven of eight earthquakes are matchedto a shared “stick-slip” design at standard branching ratio, and the eighth (Osaka, 2018) required both theindependent driver and a lowered branching ratio. A subsequent search for a physical correlate of the fittedcoupling strength (depth, magnitude, foreshock presence) across the seven shared-design events found nostatistically defensible single-variable relationship (all |r| < 0.45, p > 0.3); a superficially clean depth trend visibleat n=3 is shown to depend almost entirely on one high-leverage outlier and to weaken substantially at n=7. Wereport this as a genuine methodological contribution and an honest negative result, not a discoveredseismological law. ※ The License for codes is AGPL 3.0-or-later