# Aurora — Writing

> Engineering notes from building Aurora, a privacy-first speech workspace for iPhone. Every article reports first-party measurement or an architecture decision made while building the app.

Aurora publishes only what it is the primary source for: numbers measured on its own devices, and the reasoning behind decisions taken in its own codebase. General explainers about speech recognition are deliberately absent — better sources exist for those.

Figures in these articles are labelled where they are provisional. Where a number is measured, the article states the device, the build and the input it was measured on. Where an article describes app behaviour, it names the source file the behaviour was read from.

Each link points at the article's markdown twin, generated from the page at build; the page is the same URL without `.md`.

## Articles

- [What on-device transcription costs on an iPhone](https://auroraisyours.com/blog/on-device-transcription-cost.md): Throughput, memory ceiling, model download size and thermal behaviour for speech recognition on Apple Neural Engine. Published 2026-07-14.
- [Why Aurora does not send a bot to your meeting](https://auroraisyours.com/blog/no-meeting-bot.md): Recording from the iPhone in the room rather than joining the call as a participant; the four things that costs, and the setups device capture cannot hear — headphones during a call, and a call on the recording phone itself. The macOS app is not released and is not described here. Published 2026-08-05, updated 2026-09-18.
- [What is actually inside an Aurora diagnostic report](https://auroraisyours.com/blog/inside-a-diagnostic-report.md): The single consent switch that governs transcript, chat, recording and screenshot; the fact that it starts on for a session report; every tier; and the real members of the `.auroradiag` package. Published 2026-08-21, updated 2026-09-18.
- [What happens to your audio after you press record](https://auroraisyours.com/blog/what-happens-to-your-audio.md): The directory and format Aurora writes captured audio to, the protection attribute applied before the first sample, the crash-recovery sidecar, and every network path the app can use. Published 2026-09-18.
- [What Aurora's speaker attribution actually scores](https://auroraisyours.com/blog/speaker-attribution-measured.md): Measured word diarization error rate on LibriSpeech, Earnings-21 and AMI-SDM, the per-layer DER audit that locates the damage, and the provenance of all 37 attribution thresholds. Published 2026-09-18.
- [How Aurora decides a live word is safe to show](https://auroraisyours.com/blog/live-transcript-confirmation.md): LocalAgreement-2 with normalized token matching, a 0.6 s audio holdback before a word confirms, a dimmed unconfirmed tail, and a 2 s speaker-label dwell. Published 2026-09-18.

## Optional

- [Site-wide llms.txt](https://auroraisyours.com/llms.txt)
- [Beta status](https://auroraisyours.com/beta/index.md): What is available in beta, what is experimental, and what is still in development.
- [RSS feed](https://auroraisyours.com/feed.xml)