A clinical record and the version a reader reconstructs from it, compared line by line. Kept, compressed, dropped, inverted. Then the same record rewritten properly, for contrast.
Every line of the record carries something a reader has to act on: a warning, a dose, a timeframe, a person to call. The score is the share of that actionable weight that survives the trip into the reconstruction, weighted by how much the line matters and with an inversion counting against the total. It is a demonstration of the scoring idea, not the validated instrument.
The reconstructions here are authored examples, written to match the failure patterns that show up when an encounter is compressed: hedges deleted, negations lost, numbers dropped, warnings summarised into reassurance. Nothing is generated by a model at runtime, so the examples behave identically on every visit and on every device.
Change the audience toggle and the same record loses different things. That gap is the argument: legibility is not a property of the text, it is a property of the reader you had in mind.