Case studies

physics-lint’s rule methodology and SARIF schemas are validated against real published model checkpoints (the “borrowed credibility” methodology — see PH-SYM-001 and PH-CON-001 for the rule-level pattern). Both case studies below sit on substrates the shipped CLI and Action do not yet ingest — particle clouds and unstructured meshes — so each runs through a research harness; CLI/Action support for those substrates is planned for v1.2.0.

Summary

Case study

Substrate

Stack

Headline result

CS01 LagrangeBench

TGV2D, Dam-break, Reverse Poiseuille

JAX, particle-based GNNs (GNS, SEGNN)

Equivariance gap detected on published GNS/SEGNN checkpoints (PH-SYM-001/002 methodology) — particle substrate, run via research harness; native Action support planned for v1.2.0

CS02 PhysicsNeMo MGN

Cylinder Flow

PyTorch / PhysicsNeMo, mesh GNN (MGN)

Mass-conservation gap 5.857% GT / 5.881% MGN (PH-CON-001 structural identity, via mesh harness); PH-BC-001 structurally degenerate (D0-27)