Transcription implementation acceptance audit

Date: 2026-09-16 Status: In progress. This is a requirements audit, not a release approval.

Evidence rules

Source existence proves implementation shape, not successful execution. Fixture results prove the scenarios asserted in those fixtures. Native model results do not prove microphone ownership or remote wire compatibility. A build using the isolated Termini package does not prove the published dependency path.

Requirements and evidence

Current-source test command

The task-owned harness at ~/Library/Caches/codex-builds/hudson-transcription-contract-harness uses direct symlinks to this worktree's three source modules and their test directories. It intentionally excludes native SDK modules, which have separate checks.

swift test --package-path "$HOME/Library/Caches/codex-builds/hudson-transcription-contract-harness" \
  --scratch-path "$HOME/Library/Caches/codex-builds/hudson-transcription-contract-check"

Observed result: 41 tests across 6 suites passed. Saved output: /tmp/hudson-installation-reuse-core.log.

Next delivery checks

1. Full app rebuild passed; retain its dependency-scope limitation in delivery evidence.
2. Guide reconciliation completed by primary after both Grok flights failed. Retain the remaining acceptance requirements; do not treat fixture coverage as product acceptance.
3. Source-linked extension walkthrough added at ../examples/transcription-reference-walkthrough.md; public-API remote example is covered by the passing current-source suite.
4. Review native settings and real microphone behavior without replacing installed apps.
5. Exercise longer recording and multi-track meeting fixtures where assets and credentials allow; document external blockers precisely.
6. Reconcile the final design and evidence record before claiming goal completion.

Official model sources checked today

• Google audio transcription
• Gemini 3.8 Live
• Microsoft MAI-Transcribe-2

These pages establish documented model identity and capabilities, not live account access or successful provider requests.

Installation reuse regression check

After removing the unused installer, installation-plan API, and ModelHub module, current-source suites passed: core/cloud/ElevenLabs 41 tests, FluidAudio 8 tests, and Talkie host 32 tests. Logs: /tmp/hudson-installation-reuse-core.log, /tmp/hudson-installation-reuse-fluid.log, and /tmp/talkie-installation-reuse-host.log. The host run leaves native opt-in acceptance disabled. Existing installed models and apps were unchanged.

The complete Talkie validation app also rebuilt successfully after this cleanup (/tmp/talkie-installation-reuse-app.log). This uses the existing isolated Termini dependency fix and does not establish published-dependency acceptance. Source inspection confirms the existing Parakeet download/cancel controls remain on the same Transcription Models page below the adapter settings, routed through EngineClient to EngineService. No app launch or installation was performed.

Current native settings fixture acceptance

Rebuilt and launched only the separate acceptance bundle using fresh fixture state at talkie-transcription-settings-acceptance/installation-reuse-fixture under the user build cache. Build log: /tmp/talkie-installation-reuse-settings.log. Native accessibility interaction verified:

• Saved fixture Dictation fields and consent are loaded.
• Changing region from fixture-east to fixture-west clears consent. Save displays “Confirm audio upload before saving this remote engine.”
• Confirming fixture consent and saving displays “Engine saved and ready.”
• Activity displays both Recording: fixture-recording and Service request: fixture-request, with the duplicate-upload explanation.
• The shared Meeting provider picker opens, changes to the fixture provider, and returns to the existing-engine choice. No Meeting selection was saved.

The previously running older fixture also retained its saved fields and consent after its earlier relaunch. The observation timeout no longer blocks this fixture. This fixture contains no local adapter, so it does not verify the folder chooser or Parakeet installation controls. It uses no real keys, uploads, microphone, or installed Talkie app. Full product capture acceptance remains outstanding.

Long-file evidence: current host nativeParakeetLongFileAcceptance passed against installed Parakeet V3, with 185 seconds of generated speech. Saved output: /tmp/talkie-parakeet-long-file.log. No microphone, remote service, or app installation was used. See the testing guide for assertions and limitations.

Cloud acceptance prerequisite recheck

Rechecked the normal Gemini/Google, MAI/Azure and ElevenLabs environment variable names: none configured in this process. A noninteractive Security-framework probe of the shared, shared-dev, app and app-dev Talkie keychain services found no readable matching credentials. Gemini/Google and MAI/Azure lookups returned -25300; shared-dev ElevenLabs returned -25293, an authentication failure, not evidence that its key is absent. No secret bytes were printed and no prompt was accepted. Log: /tmp/talkie-transcription-credential-recheck.log.

The MAI client was compared again with Microsoft's current MAI-Transcribe page: model MAI-Transcribe-2, API version 2025-10-15, enhancedMode and nested modelOptions, phraseList, locales and Speech subscription-key header match the documented request. This is source-contract evidence only. Live account acceptance awaits a usable credential resolver and MAI Speech endpoint. The user has been asked for the configuration location, not secret values.

Package handoff audit

swift package dump-package succeeded for the current Hudson worktree. The core transcription target has no target dependencies. Transcription UI and Cloud each depend only on that core; FluidAudio and WhisperKit dependencies are confined to their adapter targets. ElevenLabs uses core and shared Cloud transport. All six products are registered. These are separately linked products, not separately resolved Swift packages: the root package still declares the SDK dependencies. The origin/main merge base remains 71fee5e03192dffe48056dedc0efacc94925bb62.

Full Talkie validation still depends on the isolated Termini header packaging correction. The Termini changes normalize static XCFramework headers under the GhosttyKit namespace during local installation and release packaging. They are uncommitted, including scripts/namespace-ghosttykit-headers.sh; the published Termini artifact has not been changed. Delivery must carry that fix and produce and verify the appropriate artifact before claiming normal published-dependency build acceptance. No commit, artifact publication, or installed-app replacement is authorized by these verification results.

Secret CLI discovery and live ElevenLabs acceptance

The user identified secret list as the correct store. It contains general, meetings and TTS ElevenLabs entries, but no Gemini-named entry in the current list. Previous credential checks did not cover this separate secret-cli store. At this earlier checkpoint, MAI live testing was deferred. Subsequent OpenRouter authorization and acceptance supersede that deferral.

secret run ELEVENLABS_MEETINGS_API_KEY -- ... injected the key into the opt-in elevenLabsLiveFileAcceptance test process without printing or storing its value. One Scribe v2 request through the public reference adapter passed in 1.165 seconds. The generated sentence returned “The purple lantern is on the wooden table”, eight word timing entries, and speaker_0 labels. Source digest: 06cfc5264eeaed7baeca828e2297fa722a45dac0b3eccd6cbbadc561a45a75d7. The normalized result has no provider request ID or usage values; do not infer those from the client run ID. This is short single-speaker acceptance, not a multi-speaker quality benchmark. Log: /tmp/elevenlabs-transcription-live-acceptance.log; result: /tmp/elevenlabs-transcription-acceptance-result.json.

MAI 2 via OpenRouter live wire acceptance

The user subsequently authorized MAI testing through OpenRouter. One direct request to /api/v1/audio/transcriptions with the secret-cli OPENROUTER_API_KEY returned HTTP 200 in 1.449 seconds for the same generated 2.16-second sentence used with ElevenLabs. Model: microsoft/mai-transcribe-2. Transcript: “The purple lantern is on the wooden table.” The response contained eight timed words, one segment, speaker 0 labels, language en, duration 2.16, and usage of 3 seconds at cost 0.00008333333333333333. Generation ID: gen-stt-1789588875-CPMsLUsRysv08ZU8jZSf.

Saved response: /tmp/openrouter-mai-acceptance-result.json. This supersedes the MAI-testing deferral for this OpenRouter request. It proves account access and the documented wire format, not Hudson adapter integration: the request was made by a direct acceptance script, and the OpenRouter Swift adapter remains to be implemented. No direct Azure endpoint was exercised.

MAI via OpenRouter: Swift adapter acceptance complete

On 2026-09-16, openRouterMAILiveFileAcceptance passed in 5.956 seconds through HudMAITranscriptionAdapter(route: .openRouter), using the secret-cli OPENROUTER_API_KEY and a generated 2.158-second WAV. This supersedes the earlier direct-script-only status. Talkie registers the route alongside Azure.

The normalized result contains the expected sentence, eight timed words, one speaker label, language en, the source SHA-256, operation ID, and service ID gen-stt-1789589179-jWfVPSn8Os163z1cLjrT. Reported usage is three billed audio seconds and USD 0.0000833333. This is short-file acceptance, not long-meeting, multi-speaker accuracy, or live-streaming acceptance.

Evidence: /tmp/hudson-openrouter-mai-live.log and /tmp/hudson-openrouter-mai-live-result.json.

Final OpenRouter integration checks: the shared/core/cloud/ElevenLabs harness passed 45 tests (live tests remain opt-in), Talkie host passed 33 tests, and the isolated Talkie validation application build succeeded. Both worktrees passed git diff --check. Logs: /tmp/hudson-openrouter-final-tests.log, /tmp/talkie-openrouter-host.log, and /tmp/talkie-openrouter-app.log.

Delivery remains in isolated worktrees, with the Termini binary-header packaging fix required by this validation build. Nothing was committed, pushed, published, or installed over Talkie. Gemini account acceptance remains unavailable because the checked secret-cli catalog contains no identified Gemini credential. Full-product microphone/device acceptance and long multi-speaker cloud meetings remain unverified; the completed fixtures and generated-audio runs do not prove those outcomes.

Hudson merge validation, 16 September 2026

The user authorized commit, PR, and merge after the earlier verification-only checkpoints above. The adapter branch was fast-forwarded onto origin/main 4a60eb2. The root package compiled all six new products against published Vox 9190158, without a local Vox override or a Vox source change. The default terminal-disabled package passed 72 transcription tests and 27 existing HudsonVoice tests. The experimental dependency boundary check also passed.

Commands (from the Hudson root):

swift test --scratch-path "$HOME/Library/Caches/codex-builds/hudson-transcription-merge" --filter HudsonTranscription
swift test --skip-build --scratch-path "$HOME/Library/Caches/codex-builds/hudson-transcription-merge" --filter HudsonVoiceTests
python3 scripts/apple/check-experimental-boundary.py

Logs: /tmp/hudson-transcription-merge-tests.log and /tmp/hudson-transcription-voice-regression.log. The first attempt ran out of disk while generating test debug symbols. Removing the task-owned, inactive baseline object cache allowed the same build and tests to finish successfully. Opt-in native/provider acceptance tests remain gated; these fixture totals do not represent new paid requests or physical-device acceptance. Terminal-enabled Talkie delivery still needs the separately documented Termini artifact fix.