Shared filler in synthetic transcripts makes a distiller correctly say nothing

When building synthetic session transcripts to test a pipeline that extracts what was learned, generating one distinctive narrative per fixture and wrapping it in a SHARED generic body silently invalidates the test. A good distiller judges the demonstrated work, not the asserted prose: if the transcript states a namespace-header bug and then shows edits to an unrelated timeout constant across four files including a Makefile, the honest verdict is nothing-learned, and that is what you get. Symptoms that you are measuring the fixture instead of the model: every session yields near-identical extracted struggles despite being different projects, and the only session producing output is the one whose boilerplate happens to match its narrative. Fix by making each fixture body coherent with its own scenario -- the tool calls, file paths and commands must plausibly demonstrate the specific thing the fixture claims was learned. Otherwise low yield reads as a model problem when it is a data problem.