to-recordlisted

# to-record A raw meeting export is not a record you can build on. The one this skill was designed against has no turn structure at all: a timestamp every eight seconds, speaker names in parentheses, crosstalk merged into a single line, and one speaker's verb landing four lines later inside the other speaker's reply. The export does not corrupt words. It corrupts **attribution**. This skill restructures that export into a normalised transcript and flags what is wrong with the record. It runs in four ordered steps: **ask provenance, normalise, verify, resolve flags.** Two seats run through this pipeline, and `/to-scope` defines them: **the expert** does the work and owns the process; **the builder** owns what gets built. Here only one consequence matters - **every flag is the builder's to answer**, and none is ever worth spending the expert on. ## The preservation contract **Every word between speaker markers is carried across unchanged.** Fillers, hedges, mis-transcriptions, profanity, a *"yeah and no"* that answers nothing: all of them survive verbatim. You may move words between lines and between turns. You may not add one, drop one or change one. Normalisation is **restructuring, never rewriting**. Tidying grammar, dropping filler, or resolving a vague answer into a clean one is a synthesis step wearing a parse's clothing, and it is the failure this skill exists to prevent. A model that tidies here produces a record that reads better and says something the room did n