zanmaivoicelisted

# voice Speaking is the cheapest way to get something into a vault, and the transcript is the expensive part. The words that carry the meaning are the names, and names are what speech recognition mangles. Putting them right is reading, not machinery: a garbled word resolves from the sense of the sentence the way a typo does, and most of the world's names are already known to whoever is reading. That is the default path, and for most notes it is the only one. Two things are worth knowing about the rest. **A word only the vault could know needs the vault.** A colleague's surname, a company spelled a particular way, a term nobody outside this vault uses. `zanmai.py voice lexicon` collects those and hands them to the recogniser as its initial prompt, which is the one thing that can recover a misheard name, because it is the only step that hears the audio again. It is reached for when reading does not get there, not run as a matter of course. **Rewriting a transcript has a documented failure mode: over-correction.** A model asked to fix a transcript reliably invents names that were never spoken, because a plausible name is exactly what it is good at producing. Hence the two rules that bound the reading: knowing is allowed and guessing is not, and what stays unresolved stays as it was, marked. **Who runs which leg.** A recording is a source to read, which is Reed's own ground (`reed-source-pipelines.md` covers voice memos), so Steve dispatches Reed in the background for steps