What compression actually costs

A summary is usually judged on what it contains. That is the easier test, and it is the wrong one. The interesting question is what it leaves out, because the reader will never know the difference between a detail that was never important and a detail that was silently dropped.

When a person writes a clinical note, the compression is visible in the process. They hesitate, they choose a phrase, they leave something vague on purpose. When a model writes it, the same decisions happen, but nothing in the output marks them. The prose is confident in exactly the places where a human would have signalled doubt.

Three things worth measuring

What goes missing. Not the average amount, which tells you little, but the distribution. Which categories of information survive consistently, and which disappear consistently. A summary that reliably keeps the chief complaint and reliably loses the patient's own wording is a different object from one that scatters both.

What gets added. Compression is rarely pure subtraction. Models smooth, resolve ambiguities and supply connective tissue that was not in the source. That addition is invisible unless you compare line by line against what was actually said, which is the least glamorous part of this work and the part that matters most.

Who can recover the loss. A dropped detail is survivable if the clinician knows the patient and can reconstruct it, and dangerous if the record is the only thing that travels. The same summary can be fine in one workflow and unsafe in another, which is why evaluating the text alone will keep giving us answers that do not transfer.

The failure is rarely in the sentence. It is in the handover the sentence was supposed to support.

None of this argues against summarizing. It argues for measuring the right thing. If a system saves a clinician twenty minutes and quietly removes the one phrase that would have changed a decision, the twenty minutes were not the point.