Compare two transcript contexts with local MiniLM embeddings. See paraphrased repetition that keyword counts miss, then inspect every cosine pair.
Load the bundled sample or enter 2-8 lines in each set.
The result will describe only relative semantic repetition in the selected samples.
Cohesion is the mean of every within-set cosine pair. Breadth is one minus that mean. Inspecting the pair ledger matters because an average can hide copied lines or isolated outliers.
The audit never converts language similarity into a claim about authentic preference or agent welfare. Those require different evidence.
This is evidence about selected language samples. It is not evidence of consciousness, authentic preference, welfare, financial quality, or endorsement.