Local clinical LLM benchmark seeks endorsement

This is more research signal than news, but the framing is useful: clinical NLP evals need to test data representation, not just leaderboard model swaps.

• –Running everything locally with quantized open-weight models makes the work relevant for privacy-sensitive healthcare deployment
• –FHIR serialization strategy is the interesting variable because structured clinical data often fails at the prompt boundary
• –Exact-match F1 on synthetic patients is clean but may understate real-world ambiguity, messy records, and medication reconciliation edge cases
• –No public paper or results are available yet, so the item is mainly a draft endorsement request rather than a finished benchmark release