Governing Semantic Composition to Improve LLM Output Stability

A governed classification judgment that was correct in isolation became incorrect when a neighboring output was requested from the same model on the same evidence. Reordering and routing recovered it locally but not under the full output. Correcting the governed semantic input — making explicit which definitions were authoritative and how they related, rather than leaving it for the model to infer — removed the failure under the ordinary inference path, with no change to the model and no added inference stage. Composition is semantic: inputs and requested outputs have to be governed as one system, and a semantic defect can surface as model instability.