# Domain recipe sketches

Status: design sketches only. These domains share provenance, release evidence,
container conventions, and cluster job metadata. Their training objectives and
native evaluators remain domain-specific.

## Robotics

Build an initial demonstration-to-policy recipe on
[LeRobot](https://github.com/huggingface/lerobot). Input: an identified dataset
revision, robot embodiment and calibration, observation/action schema, pretrained
policy revision, and train/test episode split. Adapt one supported policy, freeze
its preprocessing, and run a task-appropriate simulator before any robot pilot.

Report held-out task success with repeated episodes, confidence intervals,
intervention rate, and failure categories. Hold out objects, initial states, or
demonstrators deliberately; random frames from the same episode are not an
independent test split. Publish policy configuration and dataset transformations
alongside the checkpoint. The eventual recipe needs a pinned simulation/runtime
container and separate hardware execution instructions; there is no robotics
container or device support implemented here.

## Biology and medicine: three distinct paths

| Path | Starting point | Adaptation and evaluation sketch |
|---|---|---|
| Biomedical tool agent | [Biomni](https://github.com/snap-stanford/Biomni) | Preserve the tool harness; curate successful traces; evaluate executable task completion, tool correctness, and expert-reviewed evidence attribution. The agent-harness pilot can later add RL. |
| Medical text / image model | [MedGemma](https://github.com/google-health/medgemma) | Adapt using the selected model's native inputs and terms; use site/patient-separated data and external validation appropriate to the intended task. |
| Biological prediction model | [BioNeMo Recipes](https://github.com/NVIDIA-BioNeMo/bionemo-recipes) | Select one sequence, structure, or expression prediction workflow; retain native objectives and biological generalization splits. |

Choose a narrow research task before selecting a model. Specify data access and
de-identification requirements with the contributing team. Keep biomedical agent
reasoning scores separate from prediction accuracy and from evidence of clinical
utility. A benchmark improvement alone does not validate a clinical deployment.
The first deliverable should be a data manifest, one native baseline, a frozen
test protocol, and an expert-reviewed error sample, not a universal biomedical
trainer.

## Materials

Build on [MACE](https://github.com/ACEsuit/mace) for interatomic-potential
adaptation. Input: versioned structures, energies, forces and optional stress,
units, reference calculation settings, chemical coverage, and a pretrained model
revision. Fine-tune with MACE's native objective; hold out meaningful structures
or chemical systems rather than only random near-duplicate configurations.

Report energy and force error with units and normalization, stability under the
intended simulation protocol, and an out-of-distribution breakdown. If the actual
decision is discovery ranking, consider
[Matbench Discovery](https://github.com/janosh/matbench-discovery) and its specified
protocol. Do not interpret low force error as sufficient evidence of stable
molecular dynamics or useful discovery performance.

The native model, calculator settings, element coverage, data provenance, and
simulation configuration belong in the future release bundle. Materials-text
LLM adaptation would be a separate recipe and is not covered by this sketch.

## Common future artifact contract

Each domain pilot should produce an immutable input manifest, upstream/native
configuration, environment lock or image digest, checkpoint identifier, native
evaluation outputs, and a concise model card with supported use and measured
limits. A small importer can later wrap those outputs in the common evaluation
report. No domain-specific executor, GPU validation, or Hugging Face publication
is implemented by these sketches.