Member Junction
    Preparing search index...

    Builds Training Snapshots; stateless, construct once and reuse.

    Index

    Constructors

    Methods

    • The three write-once files, uploaded in parallel (independent objects) under the content-hash prefix. The keys are deterministic and the stores accept identical bytes under an existing key, so a build that failed after writing, or lost the content-hash race, leaves the same objects a later build of the same content reuses.

      The provenance file is outside the content hash (it records when each value was computed), so identical content can carry different provenance bytes; its object is keyed by its own SHA-256 so each version gets its own object. A build that loses the content-hash race can therefore leave one provenance object no snapshot references (the row references the winner's). Readers reach every file through the row's file IDs and the MJ: Files row's stored key, never by rebuilding a key.

      Parameters

      Returns Promise<StoredSnapshotFiles>

    • The series a manifest row belongs to. A forecast unit of kind grain has one row per period of a single series, so every row carries one series key: the unit's grain entity name (the unit's single series). Every other unit keeps the row's entity key (the grain record for a grain unit, the grain foreign key for a grain-x-periods unit), so a grain-x-periods forecast keeps one series per grain record, which the protocol rules refuse at bind time.

      Parameters

      Returns string

    • Whether a protocol's slice field can be written: it names an assembled feature column, or a field every given unit row carries. Training checks it before the as-of assembly (over every unit row) and the builder again when it writes the column (over the assembled rows), with this one rule.

      Parameters

      • field: string
      • columnNames: ReadonlySet<string>
      • unitRowSources: readonly SourceRow[]

      Returns boolean