protocols¶

Responsibility¶

typing.Protocol contracts each plugin kind satisfies structurally, plus runtime-checkable capability protocols the pipeline tests with isinstance instead of probing attributes (Capability protocol).

GenerativeBackend: model_id, async generate(snapshot, generation: GenerationConfig) -> GenerationResult; GenerationResult promises text.
Capabilities: SupportsKvCache (kv_cache_hits, kv_cache_reused_tokens, set_kv_cache_enabled(enabled, max_entries)), SupportsLogprobs (returns_logprobs; the transformers and api backends), SupportsAclose (async aclose()), FittableValueFunction (fit(corpus)), RequiresEmbeddings (needs_embeddings), RequiresLogprobs (needs_logprobs).
EstimatorStrategy: name, bind_snapshot, sample_masks(player_set, budget_fraction, include_minimal_masks, seed), estimate_attributions(masks, payoffs, aggregator) (Estimator); TwoPhaseEstimator: begin_allocation(masks, payoffs), sample_allocation_masks().
EstimandAggregator: estimand, coalition_weight(size, n), aggregate(masks, payoffs) (Estimand).
ValueFunction: name, score(base, candidate) (Value function).
AbsencePolicy: name, apply(snapshot, players, mask) (Absence policy); PlayerPartitioner: name, partition(snapshot) -> PlayerSet with spans (Partition).
Normalizer, AttributionRenderer, EmbeddingProvider.
protocols/plugin.py: PluginOptions (pydantic model with extra="forbid", strict types, frozen; the base class accepts no keys) and DeclaresPluginOptions (a plugin class with an options_model class variable) (Plugin options).

Imports domain types only; the pipeline and plugins packages type against these protocols and cast registry results to them.

A plugin that satisfies the protocols works without subclassing; capabilities are optional and detected with isinstance.
The orchestrator has no estimator-name switch: every estimator exposes estimate_attributions, and two-phase estimators are found by TwoPhaseEstimator.