KNOWN_LIMITATIONS.md

# Known limitations and safe claims

## What v1 represents

- Synthetic, three-dimensional, small-strain material-point histories.
- Prescribed proportional deviatoric strain paths.
- J2 plasticity with linear isotropic hardening.
- Chaboche combined hardening, explicitly marked as an experimental AgentFEM
  capability.
- A benchmark for path dependence, sequence learning and reproducible
  constitutive-surrogate studies.

## What v1 does not represent

- Structural finite-element field data.
- Experimental calibration of a named alloy.
- Fatigue life, damage initiation or crack propagation.
- Temperature dependence, creep, viscoplasticity or finite strain.
- True non-proportional multiaxial loading with rotating stress/strain
  directions.
- Validation against every commercial constitutive implementation.

## Important interpretation limits

- High random-split accuracy does not establish path or parameter
  extrapolation.
- The pointwise MLP is a history-blind scientific control, not a competitive
  constitutive architecture.
- The GRU result establishes learnability on the frozen v1 distribution; it
  does not by itself establish stable long-cycle or FE deployment.
- Material-point inference speed does not equal end-to-end structural speed-up.
- Raw stress work on plastic strain is retained as a diagnostic. For the
  Chaboche model it must not be called thermodynamic dissipation without a
  consistent split of stored and dissipated backstress energy.
- Neural-network metrics may vary slightly across hardware and library
  versions even when the physics data are deterministic.

## Changes that require a new version

- adding tensor loading paths or new constitutive models;
- changing physical parameter ranges or path-generation rules;
- changing stored variables, units or schema semantics;
- revising trajectory grouping or benchmark splits;
- correcting any physical or numerical defect in a released trajectory.

Extensions should retain v1 unchanged, publish migration notes and report
results separately on the original v1 test set and the new OOD test sets.