Establish how retrieval provenance should be presented to end users

Establish how provenance from PRISM’s population-retrieval mechanism should be presented to clinical end users without exposing individual retrieved patient records or encouraging unsupported analogy-based reasoning.

Background

PRISM supplies the LLM with mean-pooled embeddings rather than individual retrieved records, labels, or descriptive attributes. The underlying neighbors can nevertheless be enumerated offline for retrospective provenance analysis. The paper identifies the unresolved deployment question of how such provenance should be communicated to users while preserving privacy and interpretability.