docs/T2_DATA_MODEL_HANDOFF.md

# AgentFEM material-memory data and DENIM model handoff

## Public products

- Dataset: `HaomingLuo/AgentFEM-Material-Loading-Memory`
- Model: `HaomingLuo/AgentFEM-DENIM`

The dataset contains 2,660 complete material histories. A sample is always one
ordered trajectory; time points are not counted as independent samples.

| Release | Trajectories | Primary use |
|---|---:|---|
| T2 loading-memory v1 | 1,008 | loading-history baseline and proportional cyclic response |
| T2 multiaxial OOD v2 | 1,024 | ID, unseen-path and parameter-OOD comparison |
| DENIM closure v1 | 128 | incomplete internal-state closure |
| DENIM boundary v1 | 500 | path, amplitude, long-history and resolution boundaries |

These releases form one evidence chain and retain separate protocols. Pooling
all 2,660 trajectories as exchangeable supervised samples would introduce
conflicting constitutive targets and invalidate the frozen tests.

## Model product

The recommended checkpoint is `denim-expanded`, version `1.1.0`. It has 918
trainable parameters and retains a small-strain J2 skeleton with two learned
memory channels. The reference material uses three memory channels and hidden
tabulated hardening, so the task measures closure under incomplete material
state and evolution knowledge.

Published stress RMSE values are:

| Evaluation | DENIM expanded | Incomplete J2 | GRU |
|---|---:|---:|---:|
| Frozen held-out paths | 0.714 MPa | 59.541 MPa | 98.260 MPa |
| New path OOD | 0.743 MPa | 58.177 MPa | 96.959 MPa |
| Amplitude OOD | 2.294 MPa | 69.315 MPa | 95.896 MPa |
| Long history | 10.906 MPa | 54.394 MPa | 90.757 MPa |

Long-history state drift is the clearest current boundary. It should remain in
future comparisons rather than being replaced by additional ordinary-path
tests.

## Runtime package

The model repository includes an `agentfem_bundle/` directory containing:

- `model.json`: immutable architecture, state, parameter and applicability contract;
- `weights.safetensors`: runtime weights without pickle deserialization;
- `SHA256SUMS`: authenticated bundle files;
- `README.md`: offline-use notes.

The bundle is tied to exact dataset and source-model revisions. AgentFEM owns
the global Newton process, material-state commit/rollback, checkpointing and
result evidence; AgentFEM-learning owns the PyTorch provider and DENIM adapter.

## Reproducible evidence

The current package has four evidence levels:

1. dataset integrity and constitutive quality gates;
2. material-point equivalence between legacy and safe bundles;
3. path, amplitude, long-history and discretization boundary metrics;
4. serial and two-rank AgentFEM implicit execution.

The plastic automatic-differentiation tangent is currently approximate to
about 0.05–0.80% against fixed-old-state finite differences over the audited
states. The demonstrated global cases converge, but an exact consistent
plastic tangent remains an open numerical improvement.

## Material for a university research team

The public assets already provide the experimental matrix needed to organize
a computational study:

- fixed data splits and model identities;
- black-box, weak-physics, white-box and incomplete-physics comparisons;
- architecture and parameter-count evidence;
- OOD and long-history capability boundaries;
- energy, yield, incompressibility and state diagnostics;
- finite-element deployment evidence and a documented tangent limitation;
- complete generation, training and validation code.

The main method proposition available for further theoretical analysis is:

> Learn a transferable, incrementally integrable material-memory closure when
> the evolution law and internal-state description are incomplete.

Further academic work may analyze identifiability, reduced supervision,
implicit differentiation, error propagation and constitutive stability. The
published claims remain limited to the synthetic fixed-material protocols and
the verified software/runtime versions recorded above.