Member Junction
    Preparing search index...

    What a unit training run already holds when it asks for the matrix, so the assembly does not read it twice.

    interface UnitAssemblyInputs {
        Carve?: TrainingRowCarve;
        DataAccess: ICountingDataAccess;
        Metadata: IAssemblyMetadata;
        ProtocolExclusions: ProtocolExclusions;
        Resolved: ResolvedPipeline;
        SliceFields?: readonly string[];
        Unit: PredictionUnitSpec;
        UnitRead: {
            Counts: UnitResolutionCounts;
            DataCutoff: Date;
            RowEntity: AssemblyEntityInfo;
            Rows: readonly UnitRow[];
            Status: "read";
        };
    }
    Index

    Properties

    A training run's carve of the unit rows: only its kept rows are assembled, and the categorical fold counts only the development rows. Absent (an assembly outside a training run), every unit row is assembled and counted.

    ProtocolExclusions: ProtocolExclusions

    The bound protocol's exclusions (a guard field): the fields and sources that never enter the matrix, together with the pipeline's own LeakageGuard deny lists. Null when the pipeline is bound to no protocol (such a pipeline never trains; only its standalone assembly runs).

    SliceFields?: readonly string[]

    The bound protocol's slice fields. Each must be a column the assembly will write (a feature column, or one of a date column's parts) or a field of every unit row; one that is neither is refused (unresolved-column) before the as-of assembly reads anything.

    UnitRead: {
        Counts: UnitResolutionCounts;
        DataCutoff: Date;
        RowEntity: AssemblyEntityInfo;
        Rows: readonly UnitRow[];
        Status: "read";
    }