Backend capabilities

A capability is something a model backend can offer that an algorithm may need. Backends declare what they provides; algorithms declare what they requires. The rule:

An algorithm runs on a backend iff algorithm.requires <= backend.provides.

The inherited AdapterMixin.check_backend_compat enforces this and raises BackendIncompatibilityError on a mismatch. No adapter writes gate code.

Capability is a StrEnum in src/raitap/types.py. Import: from raitap.types import Capability.

Reference

Capability	Value	Status	Meaning	Provided by	Required by
AUTOGRAD	"autograd"	live	Differentiable live model with input gradients.	TorchBackend	captum gradient methods (IntegratedGradients, Saliency, GradCAM), torchattacks, foolbox, auto-LiRPA (CROWN/IBP)
TREE_MODEL	"tree_model"	live	Access to tree-ensemble structure (splits, leaf values) for TreeSHAP-style methods.	XGBoostBackend (and future sklearn / LightGBM backends)	SHAP TreeExplainer (rejected on torch/ONNX backends)
PREDICT_PROBA	"predict_proba"	live	Calibrated class-probability outputs.	XGBoostBackend	Forward pass: detects OutputKind.PROBABILITIES and skips softmax so confidences and metrics stay correct. Not a gate-requirer.

OnnxBackend provides the empty set: it runs only model-agnostic algorithms (those with empty requires, e.g. SHAP KernelExplainer, captum Occlusion / FeatureAblation).

Empty requires = model-agnostic

The default requires=frozenset() means "needs nothing special", so the algorithm runs on any backend (torch, ONNX, future tree). Set a capability only when the algorithm genuinely depends on it.

Adding a capability

Add a member to Capability only when a real algorithm needs it and a backend can provide it (ship both in the same change). A capability that nothing requires is dead and untestable.

See also

• Adding a backend declare provides on a new backend.

• Adding an algorithm declare requires on a new algorithm.