nirs4all.pipeline.dagml.native_vote module

Classification presentation evidence captured inside native scientific tasks.

The native ScoreSet remains the selection authority. These sidecars retain real fold probabilities for the library’s existing ensemble/repetition presentation; they are neither extra native score reports nor fabricated one-hot probabilities.

nirs4all.pipeline.dagml.native_vote.capture_vote_evidence(task: dict[str, Any], resolver: Any, estimator: Any, predictions: list[dict[str, Any]], train_ids: list[str], features: Callable[[list[str], bool, str], tuple[Any, dict[str, Any]]], predict: Callable[[list[str], bool, str], list[list[float]]], model_store: MutableMapping[Any, Any]) → None

Capture labelled fold/refit outputs; never fit or replace a native block.

nirs4all.pipeline.dagml.native_vote.collect_vote_evidence(model_store: dict[Any, Any]) → list[dict[str, Any]]

Extract only our typed sidecars; fitted artifact handles stay untouched.

nirs4all.pipeline.dagml.native_vote.project_vote_evidence(result: Any, records: list[dict[str, Any]], identity: Any) → None

Fill real row arrays/probabilities before the existing aggregate-twin owner.

Per-fold validation and final reports remain native. Additional fold train/ test and ensemble presentation scores are explicitly marked as library evidence, not inserted into the native ScoreSet or used for selection.