ATLAS 3.4.2 FOR macOS — TEST/BETA RELEASE AND SECURITY NOTES

STATUS

- Separate native Apple Silicon (arm64) and Intel (x86_64) DMGs.
- Requires macOS 14 Sonoma or newer.
- These TEST/Beta builds are ad-hoc signed, not Developer ID signed, and not
  Apple-notarized. macOS will show an unknown-developer warning. Follow
  ATLAS-MAC-TEST-INSTALL.txt; never disable Gatekeeper.
- Verify the downloaded DMG against SHA256SUMS.txt before opening it.

TRANSCRIPTION UPGRADE

- Captures the selected microphone as ordered 16 kHz mono PCM and durably
  flushes chunks while recording.
- Preserves the original finalized WAV and an immutable raw transcript.
- Keeps optional cleaned text separate from raw text and rejects aggressive
  cleanup changes.
- Uses local Faster-Whisper with local VAD, word timestamps, context across the
  full recording, conservative preprocessing, model fallback, and visible
  low-confidence markers.
- Generates evidence-linked summaries, decisions, action items, deadlines,
  unresolved questions, JSON, and a downloadable PDF without inventing missing
  meeting content.
- Resumes incomplete local processing after restart and prevents duplicate
  Notes, Tasks, and Calendar entries during recovery.

PRIVACY AND SECURITY

- Recording, transcription, and report generation are local by default.
- The compact bundled model is an offline emergency fallback. Higher-accuracy
  models are downloaded only when allowed, then cached locally.
- Optional Ollama cleanup remains local and is guarded by edit-distance and
  evidence checks. Ollama is not required to launch Atlas or transcribe.
- Local logs contain operational state and errors, not transcript or audio
  content.
- Atlas data is stored under:
  ~/Library/Application Support/Genesis AI/Atlas
- Apple Keychain protects supported local launcher/account secrets.

NATIVE RELEASE VALIDATION

Each architecture is built on a matching native GitHub-hosted Mac runner. The
release gate verifies the launcher, local model inference, backend routes,
window startup, binary architecture, ad-hoc signature, DMG mount/copy layout,
metadata byte count, and SHA-256 checksum before publishing the artifact.

DEPENDENCIES AND NOTICES

The app includes Faster-Whisper, CTranslate2, PyAV/FFmpeg, ONNX Runtime, NumPy,
Silero VAD, the bundled Whisper-compatible model, ReportLab, and their runtime
dependencies. License details and source links are bundled inside the app in
THIRD_PARTY_NOTICES.md.