Concept

Capture of an LLM by selection of the training corpus. The most upstream version of capture in an LLM: choosing what data the model was fitted to (what included, in what proportions, with what filtering) tilts every output regardless of how careful the runtime gate is. A captured corpus produces a captured outer message — every output is decoded by receivers using the very preconditions the corpus installed. Will be unified with related flavors in the forthcoming capture-taxonomy note.

Connections

This is a graph landing page. Use the local graph and backlinks (right) to explore, or open the interactive ontology.