# Start here: T2 multiaxial OOD research release

This package is designed for a collaborator or AI agent to continue the
research without reconstructing the project history.

## Read in this order

1. `README.md` for the complete dataset overview.
2. `artifacts/t2_multiaxial_ood_v2/QUALITY_REPORT.md` for physics validation.
3. `artifacts/t2_multiaxial_ood_v2/model_metrics.json` for all model results.
4. `models/t2_multiaxial_ood_v2/README.md` before loading checkpoints.
5. `docs/T2_MULTIAXIAL_RESEARCH_MEMO.md` for the manuscript argument and next
   method contribution.
6. `T2_V2_REPRODUCE.md` for exact commands.

## What may be claimed now

- The data generation and physical audits passed.
- History-aware models strongly outperform a history-blind control on ID data.
- Path-family OOD is substantially harder than the frozen parameter-OOD test.
- Selected hard physical constraints remove corresponding state violations.
- A state-output physics GRU did not transfer reliably to the structural gate.
- A physics-embedded neural stress integrator reduced path-OOD RMSE to 0.0237
  MPa and passed both frozen structural gates; the severe case recorded one
  trust-region fallback at complete unloading.

## What may not be claimed now

- production-ready neural constitutive modeling;
- general 3D AgentFEM learned-material integration;
- experimental-material accuracy;
- transfer of the white-box result to unknown or misspecified material laws;
- general structural convergence without the recorded global fallback;
- end-to-end acceleration in the current Python structural demonstrator.

## Highest-value continuation

Develop a gray-box evolution model in which the embedded constitutive skeleton
is deliberately incomplete and the network learns missing evolution terms or
latent internal variables. Add a verified consistent tangent, 2D/3D gates and
experimental or high-fidelity microstructural data. Retain the current splits
and structural gates unchanged so improvement is measurable.