A controlled benchmark for detecting omitted information in LLM-generated medical
summaries. Built by decomposing 460 synthetic French clinical reports into atomic facts,
deleting a known fraction of those facts at 16 levels, and summarising each degraded
report with four different LLMs. Every (report, summary) pair therefore carries a
ground-truth omission rate — the quantity an omission detector has to recover.
