What my own tests showed

How to read these

One speaker, my own Flow app. These are my readings of my own runs, reported Oct 16, 2026. They aren't Flow's error rate, and none has been checked against its recording yet. Each record says which proposal it informs.

To finish a record, paste the actual output, write what I actually said from the recording, and mark it verified.

Method

1. Set the conditions below once. Keep device, microphone, app version and cleanup settings the same for all ten runs.
2. Read each passage aloud into the Flow app. Record the audio separately (voice memo or screen recording).
3. Paste Flow's output into Actual Flow output without editing it, then lock it.
4. If you said something different from the script, mark it and write what you actually said. Judge the run against that.
5. If Flow shows a pre-cleanup transcript, paste it too. If not, leave it empty: discrepancies are then end-to-end.
6. Mark each meaning unit. Rewording, removed fillers and legitimate self-corrections are not errors.
7. After all ten, repeat each passage that needed substantive correction, to see if it recurs.
8. Then record the two unscripted brain dumps and assess them against their recordings.

This workspace never transcribes anything. Every output here is pasted from the real app.

Test conditions

New runs start with these values. Change a single run's values inside its card.