The chair holds the memory: Lessons that move between AI models, and the delivery form that decides whether bad advice is survivable: measurements from a small reproducible bench

Here is the finding in plain terms. Take one piece of advice and hand it to a small AI model two ways. Phrased as a command, the model obeys it even when it contradicts the task, and work the model does correctly on its own collapses. Phrased as an option on a menu, the identical bytes cause no harm anywhere: the model uses what fits and declines what does not. Same advice, same model, same task. The framing alone decides whether bad advice is survivable, and I measured it, on two tasks and two model families, with executable checks. This document is that measurement, the system it came from, and the record that lets you rerun it on a consumer machine. A word on the name. Diminuendo is the musical mark for growing quieter, and that is the whole job. The system does not try to catch the last error. It resolves every disharmony it can so the noise floor drops, until the one error it could not catch clears the floor and becomes audible to the person outside the room. Quiet the orchestra until the off-note is findable.