An Audit Framework for Physics-Informed Multi-Hazard Seismic Models: Gradient-Verified Coupling, Protocol- Dependent Constants, and Baseline-Anchored Operating Envelopes for Municipal Disaster Risk Reduction

Sendai Framework Target G and SDG Indicator 11.b.2 require multi-hazard risk information that local governments can write into procedures that they are able to defend, which in turn requires models whose claims can be audited. We propose an audit framework built on three questions—does the physics term reach the network, are the learned constants properties of seismology or of the protocol, and over which decisions is the score usable against which reference—and apply it to POSEIDON, a physics-informed three-head model, on 27,100 M ≥ 5 ComCat events (2010–2024) with a chronological, embargoed split and tuned tabular baselines on the same partition. Automatic differentiation shows that the coupled physics term reaches the encoder (gradient norm 50.94) while the parallel likelihood terms give exactly zero, yet the decoupled ablation reproduces the learned constants to within 0.02. Constants move by 131–317 seed standard deviations between two protocols while remaining “admissible”, which, for interval-bounded constants, is guaranteed in advance. The aftershock head has positive cost–loss value for r ∈ [0.325, 0.850], peaking at the Peirce skill score 0.441; the tsunami head has none at any r, so no life-safety decision lies inside the envelope. Every quantitative claim is stamped by the inference that supports it. The framework gives Target G’s “usable and relevant” an operational meaning that a purchasing authority can verify on ordinary hardware.