Meetings, transcribed properly. On your Mac. Nowhere else.

If we keep speaker assignment per-meeting, we sidestep cross-meeting identity entirely. You rename a speaker once, and it only affects this transcript. Identity doesn't leak across the Library.

That's the right call. Cross-meeting identity is a feature factory in disguise — every time you'd want it, the real ask is search. Ship the search index in and revisit then.

Two channels, two ramps. If we ever hit a real four-person on-mic case, we'll know — and the diarizer slot is already in the schema. Until then, the model is honest about what it can hear.