Apple first, Whisper next — first-run storyboard

Commentary Today the row says Analysing · ~2 min left, but the 1.6 GB Whisper download runs first, inside the run; the estimate is sent before the download and does not count it, and the ring fills against the clock while nothing but the download is happening. Proposed: Apple's model is already on the Mac (English came with macOS 27), so transcription starts at once, and the Whisper download starts beside it as an app-wide job in the toolbar's status zone, the sibling of the Ollama download pill. No new question, no explanation in the row, and no time estimate on a Mac that has never run anything (there is no history to base one on).

Commentary Today nothing visible has changed: the row still reads as analysis while bytes arrive, and on weak Wi-Fi it stays like this for five minutes or more. Proposed: transcription is done (about 15 s for 18 minutes of audio) and the LLM stages are running; the download continues in the pill.

Commentary Proposed: the report is readable about three minutes earlier on typical broadband (six on weak Wi-Fi). The popover is the honest progress the owner asked for: size, amount, speed and time left, in Foundation's own words, so the researcher can decide (carry on, move to faster Wi-Fi, stop). Stop Download rather than Cancel, because there is no resume: stopping throws the downloaded part away, and the note says so. The title names a speech model, not a repository; Whisper is named in the sentence. The date subtitle is the row's existing ready state (system-formatted). Today, at this moment, the row is unchanged from frame 1; its report lands near 4:20.

Commentary Apple's English model on Spanish audio gives itself low confidence (0.10–0.38 over the first two minutes, against 0.92 in the right language), so the guess is caught and Apple's text is thrown away. The run finishes with the other sessions; this one is deferred, reads Waiting for download like a queued row (information, no glyph), explains itself in the tooltip, and is transcribed by Whisper when it lands. The reverse case, English audio under a Spanish or German UI, came back as English and needs no gate. The tooltip sentence is owner-written copy.

Commentary Today the pane asks a question researchers can't answer (MLX or faster-whisper are two builds of the same engine) and shows nothing about whether the 1.6 GB model is here. Proposed: one choice stated as a behaviour (Whisper when ready, Whisper only, Apple's speech model), one sentence on when each is better, and a status row fed by the same download job as the pill: Download Now, progress with Stop Download, or On this Mac · 1.6 GB. MLX vs faster-whisper moves behind the Diagnostics preference; the interim model picker leaves the pane. The progress line is Foundation's own wording, localised by the system.

Commentary The pill has gone: absence is the signal that Whisper is ready. The new recording goes through Whisper. The first session stays Apple's transcript, because re-transcribing it would orphan any stars, tags or edits made on it; which engine wrote each session is recorded and shown in the diagnostics row only.