B.1.5:5.3 - Learned Model Pipeline

A neural-network pipeline may describe feature extraction, embedding, attention, retrieval, ranking, and explanation generation. Some blocks may be formal substrate or mechanism material, some may be constituents of a U.MethodDescription, and some may be recovered as U.Method values.

After the candidate whole represented by the pipeline and every claimed part have been independently identified as exact A.3.1 Methods, that candidate qualifies as composite only when exact methodPartOf occurrences, whole-forming claims at their A.6.RCD dispositions, accepted inputs and outputs, invariants or admissibility conditions, typed joins, fallback behavior, failure conditions, interface decisions, and reidentification rule are present. Otherwise keep the graph as a MethodDescription, mathematical lens, mechanism material, or—when an actual selection basis and receiving use exist—an A.22-selected U.Structure.

Suppose one dated training Work enacts the exact pipeline Method while three independently identified transformations occur: the feature store changes, model parameters change, and the ranking index changes. Common Work, shared data, Method order, and temporal adjacency do not establish transformation parts or one composite transformation. Without a direct transformation-composition governor, retain the three transformations and return missing-governor[transformation-composition] for the proposed three-change whole; do not call them atomic either.