One patient letter, four summarisation passes. Drag the slider and watch it come apart, then read what is left and decide whether you would know anything was missing.
Almost every study of summarisation loss measures a single compression: one encounter, one summary, one comparison. Real records are not compressed once. A ward note is summarised into a discharge letter, the letter is summarised for the patient, the patient version is summarised again for a family member on the phone, and each step is locally reasonable.
Loss compounds in a specific order. Qualifiers and hedges go first, so nothing sounds uncertain. Then quantities and timeframes, so nothing can be checked. Then warning detail, so the safety netting loses its net. Last are destinations and urgency: where to go, how fast.
Four passes is not a model run. It is an authored demonstration of that order, which is what the research is trying to establish: that compression has a predictable signature, and that the signature is invisible in the output because every pass produces something fluent.