WP19 - Adversarial Robustness and HEX8 Diagnostic¶

Decision¶

WP19 is PASS_WITH_LIMITATIONS for the bounded campaign recorded from execution source SHA dc5975b78727d9dca6d0a48b716e60f355b8799f, starting from 7f7ffbaf0b3fdda7d3ad31ba95f20a54e4719a53. The work package does not change an FEM formulation, promote HEX8R/SRI/B-bar, or change any existing maturity decision.

Adversarial robustness¶

The catalog contains 24 predeclared T1 cases: 10 positive cases and 14 expected-failure cases. The campaign produced 10 PASS and 14 EXPECTED_FAILURE_PASS, with zero FAIL and zero INVALID_EVIDENCE. All cases replayed deterministically. No NaN/Inf was accepted and the failure paths were fail-closed.

Failure class	Count
`MESH_GEOMETRY_INVALID`	4
`DUPLICATE_NODE`	1
`SINGULAR_SYSTEM`	1
`INVALID_BC_OR_LOAD`	2
`INACTIVE_DOF`	1
`INVALID_MATERIAL`	1
`UNSUPPORTED_LOAD`	1
`NONFINITE_COORDINATES`	1
`MALFORMED_GMSH`	1
`MISSING_PHYSICAL_GROUP`	1

The evidence covers inverted and near-degenerate geometry, wrong ordering, singular and inconsistent constraints, invalid loads/materials, malformed Gmsh inputs, noncontiguous IDs, rigid transforms and scale checks. It is a controlled corpus, not a universal guarantee for arbitrary meshes or input files.

HEX8 diagnostic¶

The diagnostic has nine QF rows: three axial refinement levels, three slenderness levels and three transverse-resolution levels. It uses the declared Euler cantilever tip displacement only as a diagnostic reference. The Euler relative error ranges from 38.737% to 90.650%; the axial refinement trend improves, while the slenderness sweep remains near 71% and transverse refinement changes the result from 43.243% to 38.737%.

Six same-mesh QF/C3D8 comparisons ran with CalculiX 2.20 in image qf-solver/calculix-nafems13h:2.20. The maximum full-displacement relative error is 1.997130986610937e-06 and the maximum tip error is 7.861618752234401e-07; all six cases pass the existing one-percent diagnostic displacement threshold. The deck requests displacement only, so external reaction and strain-energy comparisons are NOT_COMPARABLE.

The evidence supports LOW_ORDER_LIMITATION with a secondary MESH_DEPENDENCE diagnostic: QF and compatible C3D8 results agree while both deviate from the slender-beam Euler reference. This is compatible with a low-order bending/locking limitation, but locking is not proven by this campaign. The QF response is classified as a global-bending candidate, not as an eigenmode. No QF-specific bug was found.

No production HEX8R, selective reduced integration, B-bar or hourglass variant was evaluated or promoted. The reference TET4, TET10, HEX20 and WEDGE6 positive rows are replay evidence only and do not transfer maturity.

Golden replay and provenance¶

The WP13 golden set was run and replayed through the WP19 wrapper as a current evidence record: 8 PASS, 1 EXPECTED_FAILURE_PASS, 9/9 replay matches, and zero drift. Historical WP13 evidence remains tied to its original source SHA 94ce10a53e31ad6884383c7ec8ce1761d9533eff and is not rewritten.

Machine-readable records:

WP20 may continue with the existing J2 and external V&V scope. This record does not claim general HEX8 accuracy, universal adversarial robustness, or a new numerical formulation.