nirs4all.pipeline.dagml.general_replay module

DAG-owned inference from an already captured, trusted REFIT artifact.

This module accepts live objects only. Archive callers must verify the original artifact bytes before deserialization and retain that archive’s provenance. No training outcome or portable predictor package is invented for host models.

nirs4all.pipeline.dagml.general_replay.predict_captured_artifact(artifact: dict[str, Any], spectro: Any, *, pipeline: list[Any], target_names: list[str] | None = None) → tuple[ndarray, dict[str, Any]][source]

Predict every input row through a native PREDICT-only execution plan.

artifact contains the fitted estimator (including its X-chain) and optional y_transform captured at REFIT. pipeline is the corresponding concrete training topology, including an optional training-only splitter. Target names come from training metadata, defaulting to ['y'] for one target. Input labels are never required or used: this is inference, not a new validation experiment. Returned arrays follow input storage order; evidence retains the unmodified native identity-keyed execution results.