Build report

What was built overnight, and what each piece was actually tested against.

HQ / Recorder / Build report

Stages 1 to 4 are built and checked on your own board recording.

The engine runs on the Mac's graphics chip, the whole chain works with the network switched off, an hour of meeting costs about 23.5 minutes of the Mac's time, voices are told apart, and a file dropped into the folder transcribes itself with nobody watching.

Where everything lives

your ruling, 2 Aug

Everything sits in Claude HQ / Recorder, Google Drive included.

Overnight the audio and transcripts had been kept in a Mac-only folder, on the reading that nothing should reach Google. You ruled that your Drive counts as safe as the Mac, so it all moved back into the project folder: recordings, transcripts, models and logs. The transcription itself still runs entirely on this Mac with no account and no internet, and these pages sit behind your Google sign-in. Claude writes a summary only when you ask for one, meeting by meeting.

The checks

each one could have failed

Stage	Check	Result	What happened
Stage 1	Saleh's real 2h12m board recording transcribes	passed	9,668 words in 1,248 lines, languages ar, en
Stage 1	runs on the M4 graphics chip, not the processor	passed	ggml_metal_device_init: GPU name: MTL0 (Apple M4) · load_backend: loaded MTL backend from /opt/homebrew/Cellar/ggml/0.18.0/libexec/libggml-metal.so
Stage 1	network sandbox actually blocks the network	passed	blocked inside the sandbox (curl exit 6); the same request outside it works
Stage 1	full pipeline runs with the network denied	passed	transcribed twovoice.wav in 13s with no network; 79 words out
Stage 2	speed measured on a real meeting, not estimated	passed	15.4 minutes of work per hour of meeting (3.89x real time) on the Mac mini M4; the blueprint had guessed 10 to 20 minutes
Stage 3	two-voice clip (Arabic voice + English voice) splits correctly	passed	voice A → label 0, voice B → label 1
Stage 3	voice separation scored against a clip where we know the truth	passed	campplus_zh_en 0.5 2 99.1% 99.8% 2.6
Stage 4	watched folder is armed	passed	launchd job loaded: - 0 sa.alfadhel.recorder
Stage 4	drop a file in, walk away, transcript appears by itself	passed	appeared on its own after 20s, no clicks

Where the plan changed

and why

Voice separation uses a different door to the same modelThe blueprint named pyannote. Its official version needs an account and an accepted licence, which your rules forbid. The same model, converted and published as an ordinary file on GitHub with no sign-up, is what runs here. We do our own tidying afterwards, which the hosted version would have done for us.

The recording is cut into pieces before transcribingYour meetings switch between Arabic and English. The engine picks one language per run, so one pass over two hours would have labelled the whole thing Arabic and mangled the English. It is cut at natural pauses instead, and each piece gets its own language decision.

The offline proof is a sandbox, not the WiFi switchPulling the network would also cut the link you reach this Mac through while you sleep. Instead the run is put in a cage where every network call is refused, which is a stricter test: it fails even on a local connection.

Version history

newest first

Version	When	What changed
1	2 Aug	First build report: stages 1 to 4, checks, and the Drive privacy fix