Model Cards and Dataset Datasheets

Every deployable model ships a card stating its purpose, validated envelope, metrics, and known limits — an honest, machine-readable statement of what it can and cannot do.

Documenting what a model is for

A model card is a structured document that travels with a model artifact, stating in plain terms what the model does, the machine and subsystem it serves, the operating envelope it was validated on, its measured performance and calibration, and its known limitations. It is generated from the tracked run and lineage, not written from memory, so it cannot drift from the artifact.

Cards are honest by mandate. A breeder equilibrium surrogate's card states the exact regime it was validated on and where it extrapolates. A burner model's card states the honest gates plainly: that the plug is over-stressed 3 to 3.9x at the design bore, that the operating regime is 166 to 830x beyond any device and cannot be post-dicted today, that He-3 demand is about 400x domestic supply per commercial unit, and that availability is 0.86 to 0.995 against a hyperscale requirement of 0.99982. A card never implies a capability the physics does not support.

Card contents

• Purpose, machine, subsystem, and intended use
• Validated operating envelope and out-of-envelope behavior
• Metrics: accuracy, calibration, latency, robustness
• Training data datasheet reference and lineage hashes
• Known limitations and honest gates where applicable

card = generate_card(model, run, lineage)
assert card.envelope   is not None
assert card.limitations is not None      # no empty limits section
assert card.metrics.calibration_ece <= TH.ece
# burner cards must reference honest_gates registry entry
if model.machine == 'burner':
    assert 'honest_gates' in card.references

Dataset datasheets are the data-side analogue: they state a dataset's composition, collection conditions, known biases, and curation decisions. Together, cards and datasheets make a model self-describing to reviewers and to the AI systems that consume this repository, and they are a required output of the validation gates.