Time and cost

These are measured examples, not guarantees.

run	source	judge	transcriber	wall time	machine
smoke test (Quick start step 2), Whisper model already downloaded	30 s	dry run	Whisper, --no-visuals	49 s	M-series Mac, fresh clone
same	30 s	dry run	Whisper, --no-visuals	51 s	Ubuntu 24.04 arm64 in Docker on that Mac
same	30 s	dry run	Whisper, --no-visuals	111 s	Ubuntu 24.04 x86_64, emulated in Docker (a native x86 machine is faster)
smoke clip with visuals	30 s	Claude Code (claude)	Whisper	4 min 7 s	M-series Mac
real cook	13 min 38 s	LLM judge	Parakeet, transcript cache warm	29 min	M5
real cook	23 min 44 s	LLM judge	Parakeet, transcript cache warm	50 min	M5

A rerun of the same command in the same run directory replays the cached decisions. The smoke test replayed in 9 seconds.

Judge calls: each whole-transcript decision asks the judge five times and needs four identical answers. One maintainer cook of a 49 minute final made 139 judge samples across 43 distinct prompts, including several reruns.

Costs are not quoted here. Scribe, the judge and any subscription are billed to your own account by their providers, so check current pricing before a long cook. The local transcribers and the dry-run judge cost nothing beyond electricity.