Skip to content

Fastest and only macOS Dictation app with on-device STT and custom trained AI enhancement model. Windows, iOS testflight! Android, Linux waitlist open! A local Wispr Flow alternative. Follow https://x.com/fluidvoiceapp for more

docs: define meeting transcription milestones

altic-dev356fae66bdbf
diff --git a/MEETING_TRANSCRIPTION_PRD.md b/MEETING_TRANSCRIPTION_PRD.md
new file mode 100644
index 00000000..e452c006
--- /dev/null
+++ b/MEETING_TRANSCRIPTION_PRD.md
@@ -0,0 +1,808 @@
+# FluidVoice Meeting Transcription PRD
+
+**Status:** Draft for product alignment
+**Last updated:** 2026-08-05
+**Target:** FluidVoice for macOS 15+
+**Initial release:** English, local recording, offline transcription after Stop
+**Long-term direction:** Optional meeting detection, remembered speaker identities, and live progressive transcription
+
+---
+
+## 1. How to use this document
+
+Every tracked requirement has a checkbox and one of these tags:
+
+- **[MUST · FOUNDATION]** — Must be designed correctly before implementation because later milestones depend on it.
+- **[MUST · V1]** — Release blocker for the first trustworthy offline meeting-transcription release.
+- **[OPTIONAL · V1]** — Valuable, but may move behind the first release without invalidating the core experience.
+- **[MUST · Mx]** — Required for that named milestone; M1 is the first working vertical slice, M2 is the trustworthy V1 gate, and M3–M5 are staged follow-ups.
+- **[OPTIONAL · Mx]** — Valuable within that milestone but safe to defer.
+- **[CONDITIONAL MUST · Mx]** — Mandatory only if its parent optional capability is enabled.
+- **[OPTIONAL · SECURITY]** — Security hardening that must not be claimed until implemented and verified.
+- **[FUTURE · LIVE]** — Explicitly outside V1; the V1 architecture must leave a clean path for it.
+
+Checkboxes should be checked only when the requirement and its acceptance criteria are complete. A compiling implementation alone is not completion.
+
+### Document map
+
+- [Agreed product requirements](#3-agreed-product-requirements--implementation-status)
+- [Goals and non-goals](#4-goals-and-non-goals)
+- [High-level milestones](#5-high-level-milestones)
+- [User experience and canvas states](#7-user-experience-and-canvas-states)
+- [Session lifecycle](#8-session-lifecycle)
+- [Target architecture](#9-target-architecture)
+- [Audio capture](#11-audio-capture-design)
+- [Offline transcription](#12-offline-english-transcription-pipeline)
+- [Speaker identity](#13-speaker-identity-and-correction)
+- [Supported platforms](#14-supported-meeting-platforms)
+- [Meeting detection](#15-meeting-detection-and-suggestion-prompt)
+- [Menu-bar lifecycle](#16-menu-bar-lifecycle-and-notifications)
+- [Privacy and retention](#17-permissions-consent-privacy-and-retention)
+- [Acceptance gates](#21-acceptance-and-release-gates)
+- [Open decisions](#22-open-decisions-for-alignment)
+
+---
+
+## 2. Product summary
+
+FluidVoice should let someone record an online or in-room meeting locally, stop recording, and receive an English transcript organized by speaker. Online meetings should preserve the user's microphone separately from meeting-app audio so clean speech from a confirmed personal microphone can be labeled **You** while remote participants are diarized. Shared microphones, leakage, and ambiguous overlap must remain explicitly unknown.
+
+The initial release deliberately records first and transcribes after Stop. This protects recording reliability and resource use. The underlying session and transcript models must still support provisional, revisable segments so a future live mode can progressively fill the same transcript canvas without a rewrite.
+
+Speaker diarization answers **who spoke when**, but it does not inherently know real names. FluidVoice may propose a name using visible meeting UI and confirmed local voice profiles, but the user remains the authority through a compact confirmation:
+
+```text
+Speaker 2 · Was this Erik? [Yes] [No]
+```
+
+---
+
+## 3. Agreed product requirements — implementation status
+
+The direction below is agreed. Its checkboxes remain open until the corresponding behavior is implemented and verified.
+
+- [ ] **[MUST · V1] DEC-001 — English-only initial release.**
+ - Persist `languageCode = "en"` in the session, processing request, history, and exports.
+ - Do not scatter hard-coded English assumptions through capture or transcript models.
+
+- [ ] **[MUST · V1] DEC-002 — Local-first data path.**
+ - V1 capture and processing never upload meeting audio, transcripts, participant names, or voice embeddings.
+ - Network activity is limited to model downloads and exports explicitly initiated by the user.
+ - Analytics and diagnostics must exclude their contents.
+
+- [ ] **[MUST · V1] DEC-003 — Explicit recording action.**
+ - FluidVoice never begins recording automatically.
+ - A meeting-detection prompt may suggest recording, but the user must click **Record**.
+
+- [ ] **[MUST · V1] DEC-004 — Offline-first processing.**
+ - V1 records during the meeting and begins ASR/diarization only after Stop.
+ - The recording path remains useful even when transcription models are not yet ready.
+
+- [ ] **[MUST · FOUNDATION] DEC-005 — Live-ready schema.**
+ - Segments have stable IDs, revisions, and provisional/final state from the start.
+
+- [ ] **[MUST · V1] DEC-006 — Separate online-call tracks.**
+ - Meeting/application audio and microphone audio remain separate and timestamped.
+ - They are never irreversibly mixed before transcription and identity processing.
+
+- [ ] **[MUST · V1] DEC-007 — Two capture modes.**
+ - **Online call:** selected application audio plus selected microphone.
+ - **In-room meeting:** selected microphone only; the microphone is not automatically labeled **You**.
+
+- [ ] **[MUST · V1] DEC-008 — Persistent global state.**
+ - Recording and processing are app-level activities, not owned by the sidebar tab or main window.
+
+- [ ] **[MUST · V1] DEC-009 — Honest hardware capability.**
+ - Apple Silicon receives speaker diarization and identity features.
+ - Intel may record and produce a plain transcript, but must not promise remote-speaker separation until supported.
+
+- [ ] **[MUST · V1] DEC-010 — Honest platform claims.**
+ - Zoom, Google Meet, and Microsoft Teams are planned optimized targets, not marketed as optimized until their acceptance matrix passes.
+
+---
+
+## 4. Goals and non-goals
+
+### 4.1 Goals
+
+- [ ] **[MUST · V1] GOAL-001 — Trustworthy capture.** Preserve usable meeting audio across long sessions, device changes, window closure, and recoverable failures.
+- [ ] **[MUST · V1] GOAL-002 — Useful offline transcript.** Produce a timestamped English transcript with truthful speaker attribution and clear fallbacks.
+- [ ] **[MUST · V1] GOAL-003 — Clear local identity.** Label clean speech from a confirmed personal online-call microphone as the configured local user or **You**; preserve uncertainty otherwise.
+- [ ] **[MUST · V1] GOAL-004 — Correctable speaker model.** Support global rename, individual reassignment, merge, and undo.
+- [ ] **[MUST · M3] GOAL-005 — Reusable identities.** Suggest previously confirmed speakers across meetings without silently assigning names.
+- [ ] **[MUST · M4] GOAL-006 — Helpful meeting awareness.** Suggest recording at the right time without stealing focus or creating false confidence.
+- [ ] **[FUTURE · LIVE] GOAL-007 — Progressive transcript.** Fill the same canvas during a meeting and reconcile it after Stop.
+
+### 4.2 Non-goals for V1
+
+- [ ] **[MUST · V1] NONGOAL-001 — No automatic recording.**
+- [ ] **[MUST · V1] NONGOAL-002 — No live transcription requirement.**
+- [ ] **[MUST · V1] NONGOAL-003 — No automatic summaries, action items, calendar integration, or meeting bots.**
+- [ ] **[MUST · V1] NONGOAL-004 — No claim of legal compliance or consent determination.** Recording laws and workplace policies vary.
+- [ ] **[MUST · V1] NONGOAL-005 — No promise of perfect name identification.** Voice and visual matching produce candidates, not facts.
+- [ ] **[MUST · V1] NONGOAL-006 — No tab-specific browser audio claim.** ScreenCaptureKit audio filtering is application-level.
+- [ ] **[MUST · V1] NONGOAL-007 — No echo-cancellation claim.** Speaker leakage may be detected or reduced later, but is not solved in V1.
+- [ ] **[MUST · V1] NONGOAL-008 — No reuse of dictation capture as the meeting recorder.** Meeting capture has different lifecycle, durability, and ownership requirements.
+
+---
+
+## 5. High-level milestones
+
+### Milestone 0 — Decisions and durable foundations
+
+- [ ] **[MUST · FOUNDATION] M0-EXIT-001** Resolve FOUNDATION and V1 open decisions in Section 22; M3/M4 decisions remain gated by their own milestones.
+- [ ] **[MUST · FOUNDATION] M0-EXIT-002** Approve the session, track, speaker, identity, and transcript models.
+- [ ] **[MUST · FOUNDATION] M0-EXIT-003** Approve the capture/recovery format and app-wide state machine.
+- [ ] **[MUST · FOUNDATION] M0-EXIT-004** Approve privacy, retention, platform-support, and Intel fallback copy.
+
+### Milestone 1 — Working vertical slice (not yet a release)
+
+Goal: prove the complete path before polishing secondary automation.
+
+- [ ] **[MUST · M1] M1-EXIT-001** A new Meeting Transcription tab can manually start an online or in-room recording.
+- [ ] **[MUST · M1] M1-EXIT-002** Online mode writes separate system/application and microphone tracks.
+- [ ] **[MUST · M1] M1-EXIT-003** Stop ends capture, runs offline English transcription, and displays a timestamped result.
+- [ ] **[MUST · M1] M1-EXIT-004** Apple Silicon output includes remote speaker labels and, for a confirmed personal microphone with clean speech, a **You** track; uncertain microphone speech remains unknown.
+- [ ] **[MUST · M1] M1-EXIT-005** The menu-bar menu shows **Stop Meeting Recording** first while capture is active.
+- [ ] **[MUST · M1] M1-EXIT-006** Window closure and sidebar navigation do not stop the session.
+
+### Milestone 2 — Trustworthy offline V1
+
+Goal: make the working slice safe to ship.
+
+- [ ] **[MUST · V1] M2-EXIT-001** Recoverable chunking, atomic manifests, source-health monitoring, and relaunch recovery pass.
+- [ ] **[MUST · V1] M2-EXIT-002** Transcript correction, history, export, deletion, and retention behavior pass.
+- [ ] **[MUST · V1] M2-EXIT-003** Permissions, menu-bar lifecycle, notifications, accessibility, and multi-display behavior pass.
+- [ ] **[MUST · V1] M2-EXIT-004** Zoom, Meet, Teams, generic-source, headphone, speaker, device-change, and one-hour capture matrices pass at their claimed tiers.
+- [ ] **[MUST · V1] M2-EXIT-005** Apple Silicon and Intel capabilities are accurately gated and described.
+
+### Milestone 3 — Remembered speaker identities
+
+- [ ] **[MUST · M3] M3-EXIT-001** Explicit opt-in voice profiles store only confirmed clean samples.
+- [ ] **[MUST · M3] M3-EXIT-002** Future meetings produce conservative name candidates with **Yes/No** confirmation.
+- [ ] **[MUST · M3] M3-EXIT-003** Rename, unlink, forget-one-person, and delete-all-profile controls pass.
+
+### Milestone 4 — Platform intelligence and meeting suggestions
+
+- [ ] **[MUST · M4] M4-EXIT-001** Versioned platform adapters power detection and capability reporting.
+- [ ] **[MUST · M4] M4-EXIT-002** Opt-in multi-signal meeting detection passes false-positive and cooldown checks.
+- [ ] **[MUST · M4] M4-EXIT-003** Notch and top-center prompts remain nonactivating and accessible.
+- [ ] **[OPTIONAL · M4] M4-EXIT-004** In-memory active-speaker visuals improve name candidates without affecting recording reliability.
+
+### Milestone 5 — Live progressive transcription
+
+- [ ] **[FUTURE · LIVE] M5-EXIT-001** Bounded live processing emits provisional segments while durable recording remains independent.
+- [ ] **[FUTURE · LIVE] M5-EXIT-002** The canvas remains stable during revisions, selection, and scrolling.
+- [ ] **[FUTURE · LIVE] M5-EXIT-003** Stop performs final offline reconciliation without losing user corrections.
+
+---
+
+## 6. Information architecture
+
+- [ ] **[MUST · V1] IA-001 — Sidebar placement.** Add **Meeting Transcription** directly below **File Transcription** in the existing **Use** group.
+- [ ] **[MUST · V1] IA-002 — Separate product concepts.** File Transcription imports existing media; Meeting Transcription records and owns meeting sessions.
+- [ ] **[MUST · FOUNDATION] IA-003 — Shared transcript domain.** Imported files and captured meetings should eventually use the same stable transcript/speaker model even if their orchestration remains separate.
+- [ ] **[MUST · V1] IA-004 — One state-driven meeting canvas.** The tab renders setup, recording, processing, result, interrupted, and recovery states from the app-level coordinator.
+- [ ] **[MUST · V1] IA-005 — Recent meetings.** Recent sessions appear in the Meeting tab and reopen the same canvas.
+- [ ] **[MUST · V1] IA-006 — Stable window behavior.** Changing tabs, minimizing, or closing the main window does not stop recording or processing.
+- [ ] **[MUST · V1] IA-007 — Mac command access.** Stop, open meeting, copy, find, export, and settings actions are reachable by keyboard and appropriate menus.
+
+---
+
+## 7. User experience and canvas states
+
+### 7.1 Idle/setup
+
+```text
+Meeting Transcription Stored on this Mac
+
+[ Online call ] [ In-room meeting ]
+
+Meeting audio Zoom / Google Chrome / Choose application… Ready
+Microphone MacBook Microphone Ready
+Language English
+Speaker separation Automatic on Apple Silicon
+
+Headphones recommended for the cleanest speaker separation.
+
+ [ Start recording ]
+
+Recent meetings
+Design sync · Today · 42 min · 4 speakers
+```
+
+- [ ] **[MUST · V1] UX-SETUP-001** Show mode, meeting source, microphone, English, permission status, model readiness, and storage readiness.
+- [ ] **[MUST · V1] UX-SETUP-002** Show separate meeting-audio and microphone meters before Start when sources can be sampled safely.
+- [ ] **[MUST · V1] UX-SETUP-003** Recommend headphones without blocking speaker use.
+- [ ] **[MUST · V1] UX-SETUP-004** Default the editable title to platform plus date/time.
+- [ ] **[MUST · V1] UX-SETUP-005** Disable Start with a plain explanation when required permission, source, or disk readiness is missing.
+- [ ] **[MUST · V1] UX-SETUP-006** Model unavailability must not prevent recording; it must clearly defer processing until models are available.
+- [ ] **[OPTIONAL · V1] UX-SETUP-007** Allow an expected in-room speaker count.
+- [ ] **[MUST · V1] UX-SETUP-008** Show an Apple Silicon/Intel capability explanation before recording, not only after processing.
+
+### 7.2 Recording
+
+```text
+● Recording · 24:18 Zoom · MacBook Microphone
+
+Meeting audio ▂▅▃▆ Healthy
+Microphone ▃▂▅▃ Healthy
+
+The transcript will appear after the meeting.
+
+ [ Stop & Transcribe ]
+```
+
+- [ ] **[MUST · V1] UX-REC-001** Show elapsed time, meeting source, microphone source, and independent track health.
+- [ ] **[MUST · V1] UX-REC-002** Make **Stop & Transcribe** the primary action.
+- [ ] **[MUST · V1] UX-REC-003** Stop immediately without confirmation because it preserves the session and begins processing.
+- [ ] **[MUST · V1] UX-REC-004** Put **Discard Recording** behind a destructive confirmation and keep it visually secondary.
+- [ ] **[MUST · V1] UX-REC-005** Show recording state in the tab, sidebar, and menu bar without relying on color alone.
+- [ ] **[MUST · V1] UX-REC-006** Warn immediately when a source disappears and preserve the healthy source.
+- [ ] **[OPTIONAL · V1] UX-REC-007** Pause/resume. If deferred, omit it rather than creating a misleading pseudo-pause.
+
+### 7.3 Processing
+
+- [ ] **[MUST · V1] UX-PROC-001** Present explicit stages: **Saving → Identifying speakers → Transcribing → Finalizing**.
+- [ ] **[MUST · V1] UX-PROC-002** Remove the recording indicator as soon as capture actually stops; use a visually distinct processing state.
+- [ ] **[MUST · V1] UX-PROC-003** Continue processing when the main window closes.
+- [ ] **[MUST · V1] UX-PROC-004** Resume safely at a persisted phase/chunk boundary after relaunch.
+- [ ] **[MUST · V1] UX-PROC-005** Preserve audio on failure and offer **Retry**, **Reveal Recording**, and **Export Audio**.
+- [ ] **[MUST · V1] UX-PROC-006** Preserve a partial transcript when useful and label it incomplete.
+- [ ] **[OPTIONAL · V1] UX-PROC-007** Populate finalized transcript sections before the entire offline job completes.
+
+### 7.4 Result
+
+```text
+Design sync 42:16
+[ Barath (You) ] [ Speaker 1 · Rename ] [ Speaker 2 · Rename ]
+
+00:04 Barath (You)
+ Let's start with the onboarding changes.
+
+00:11 Speaker 1
+ I pushed the updated flow yesterday.
+
+[ Copy ] [ Export ] [ Reveal Recording ] [ New Meeting ]
+```
+
+- [ ] **[MUST · V1] UX-RESULT-001** Show timestamped, speaker-grouped, selectable transcript text.
+- [ ] **[MUST · V1] UX-RESULT-002** Global speaker rename updates every segment referencing that speaker ID.
+- [ ] **[MUST · V1] UX-RESULT-003** Support Copy, Find, text export, JSON export, reveal recording, delete audio, and new meeting.
+- [ ] **[MUST · V1] UX-RESULT-004** Deleting audio may preserve the transcript.
+- [ ] **[MUST · V1] UX-RESULT-005** Allow individual segment reassignment.
+- [ ] **[MUST · V1] UX-RESULT-006** Allow merging mistaken duplicate speakers.
+- [ ] **[OPTIONAL · V1] UX-RESULT-007** Split one wrongly combined speaker into two.
+- [ ] **[MUST · V1] UX-RESULT-008** Support undo for speaker rename, reassignment, merge, and identity confirmation.
+- [ ] **[MUST · V1] UX-RESULT-009** **Export Audio…** offers authoritative separate tracks and an optional mixed convenience file generated on demand; it never silently replaces the originals.
+- [ ] **[MUST · V1] UX-RESULT-010** Text/JSON exports omit embeddings, model fingerprints, internal confidence vectors, and private profile metadata.
+
+---
+
+## 8. Session lifecycle
+
+```mermaid
+stateDiagram-v2
+ [*] --> Idle
+ Idle --> Preparing: Start or Set up
+ Preparing --> Recording: Permissions and sources ready
+ Preparing --> Idle: Cancel or failure
+ Recording --> Stopping: Stop and transcribe
+ Recording --> RecordingDegraded: One source becomes unhealthy
+ RecordingDegraded --> Recording: Source recovered on a new chunk boundary
+ RecordingDegraded --> Stopping: Stop and transcribe available tracks
+ Recording --> Interrupted: Capture-wide stoppage
+ Stopping --> Processing: Durable chunks finalized
+ Interrupted --> Processing: Recover available audio
+ Interrupted --> Recording: Safe resume when supported
+ Processing --> Completed
+ Processing --> ProcessingFailed
+ ProcessingFailed --> Processing: Retry
+ Completed --> Idle: New meeting
+```
+
+- [ ] **[MUST · FOUNDATION] STATE-001** Implement an app-wide state machine: `idle`, `preparing`, `recording`, `recordingDegraded`, `stopping`, `processing`, `completed`, `interrupted`, `failed`.
+- [ ] **[MUST · FOUNDATION] STATE-002** Persist every durable state transition atomically.
+- [ ] **[MUST · V1] STATE-003** Treat source loss, app exit, permission revocation, disk exhaustion, sleep/wake, and clock discontinuity as typed events rather than one generic error.
+- [ ] **[MUST · V1] STATE-004** Keep capture and processing failure domains separate.
+- [ ] **[MUST · V1] STATE-005** Never delete the source recording automatically because transcription failed.
+- [ ] **[MUST · V1] STATE-006** Define termination behavior: closing the window continues; quitting presents a clear choice and never implies background survival after process exit. **Stop & Quit** persists `processingPending` and says transcription resumes next launch.
+- [ ] **[MUST · V1] STATE-007** Allow only one active meeting capture session at a time.
+- [ ] **[MUST · V1] STATE-008** If an older meeting is processing when a new recording starts, checkpoint and queue/pause processing so durable capture receives priority; resume processing after recording stops.
+
+---
+
+## 9. Target architecture
+
+```mermaid
+flowchart TD
+ UI[Meeting Tab / Menu Bar / Prompt] --> COORD[MeetingSessionCoordinator]
+ COORD --> ARB[Audio Activity Arbiter]
+ COORD --> CAP[MeetingCaptureEngine]
+ CAP --> SYS[Application Audio Track]
+ CAP --> MIC[Microphone Track]
+ SYS --> STORE[Rolling Chunks + Atomic Manifest]
+ MIC --> STORE
+ STORE --> PIPE[Offline Meeting Pipeline]
+ PIPE --> DIAR[Remote Diarization]
+ PIPE --> ASR[English ASR]
+ DIAR --> ID[Speaker Identity Resolver]
+ ASR --> TIMELINE[Timestamped Transcript Timeline]
+ ID --> TIMELINE
+ TIMELINE --> SESSION[MeetingSessionStore]
+ SESSION --> UI
+```
+
+### 9.1 Required components
+
+- [ ] **[MUST · FOUNDATION] ARCH-001 — `MeetingSessionCoordinator`.** App-wide `@MainActor` owner exposed through `AppServices`; coordinates UI state without owning real-time audio callbacks.
+- [ ] **[MUST · FOUNDATION] ARCH-002 — `MeetingCaptureEngine`.** Actor/queue-backed ScreenCaptureKit capture owner with bounded callback work.
+- [ ] **[MUST · FOUNDATION] ARCH-003 — `MeetingSessionStore`.** Disk-backed, versioned meeting/session index; do not extend the file-transcription UserDefaults store.
+- [ ] **[MUST · FOUNDATION] ARCH-004 — `MeetingProcessingPipeline`.** Resumable offline orchestration for track preparation, diarization, ASR, timeline merge, and finalization.
+- [ ] **[MUST · FOUNDATION] ARCH-005 — `MeetingPlatformRegistry`.** Local registry of platform detection/capture/name capabilities and caveats.
+- [ ] **[MUST · M3] ARCH-006 — `SpeakerProfileStore`.** Versioned local person/profile database independent of meeting deletion.
+- [ ] **[MUST · M4] ARCH-007 — `MeetingDetectionService`.** Opt-in sustained multi-signal detector.
+- [ ] **[MUST · M4] ARCH-008 — `MeetingSuggestionPresenter`.** Separate from the latency-sensitive dictation notch/bottom-overlay state machine.
+- [ ] **[MUST · FOUNDATION] ARCH-009 — Audio activity arbiter.** Explicitly owns conflicts among dictation, command/edit capture, dictionary recording, and meeting capture.
+
+### 9.2 Current code integration
+
+- [ ] **[MUST · FOUNDATION] CODE-001** Add the meeting coordinator lazily to `AppServices`; extend termination shutdown to flush capture without relying on graceful exit for recovery.
+- [ ] **[MUST · V1] CODE-002** Add a distinct sidebar destination and menu-bar navigation destination for Meeting Transcription.
+- [ ] **[MUST · FOUNDATION] CODE-003** Rename the current file-oriented `MeetingTranscriptionView`/service concept to File Transcription before adding the real meeting view, avoiding two meanings for “meeting transcription.”
+- [ ] **[MUST · FOUNDATION] CODE-004** Replace rename-unstable segment IDs. The current file segment ID derives from speaker text and timestamp.
+- [ ] **[MUST · FOUNDATION] CODE-005** Return rich diarization output instead of discarding FluidAudio cluster embeddings and quality metadata.
+- [ ] **[MUST · V1] CODE-006** Keep meeting sessions out of `FileTranscriptionHistoryStore`; it uses UserDefaults, caps entries, and lacks recovery/audio state.
+- [ ] **[MUST · V1] CODE-007** Add an appropriate Screen & System Audio usage description and explicit permission-state handling.
+- [ ] **[MUST · V1] CODE-008** Extend the existing `MenuBarManager` rather than creating a second status item.
+- [ ] **[MUST · V1] CODE-009** Do not call `ASRService.start()` for meeting recording; reuse only safe file/provider APIs during offline processing.
+- [ ] **[MUST · V1] CODE-010** Ensure existing view disappearance and ASR shutdown paths cannot stop the meeting coordinator.
+
+---
+
+## 10. Meeting session and data model
+
+### 10.1 Meeting session
+
+- [ ] **[MUST · FOUNDATION] MODEL-001** Define a versioned `MeetingSession` containing:
+ - schema version and session UUID;
+ - title and `languageCode`;
+ - online/in-room capture mode;
+ - platform profile and captured application identity;
+ - selected microphone identity;
+ - started/ended timestamps and monotonic timebase metadata;
+ - session state and typed interruption/failure history;
+ - audio track manifests;
+ - session speakers and transcript segment IDs;
+ - retention/deletion state;
+ - processing checkpoints.
+
+### 10.2 Audio tracks and chunks
+
+- [ ] **[MUST · FOUNDATION] MODEL-002** Define `MeetingAudioTrack` with stable ID, kind (`applicationAudio` or `microphone`), source/device metadata, format, shared timebase mapping, health, and ordered chunks.
+- [ ] **[MUST · FOUNDATION] MODEL-003** Define `MeetingAudioChunk` with stable ID, sequence, file URL, presentation start/end, discontinuity metadata, checksum, and finalization state.
+- [ ] **[MUST · V1] MODEL-004** Make every finalized chunk independently recoverable.
+
+### 10.3 Speakers and transcript segments
+
+- [ ] **[MUST · FOUNDATION] MODEL-005** Define stable `SessionSpeakerID` separately from editable display name and source diarization cluster.
+- [ ] **[MUST · FOUNDATION] MODEL-006** Define transcript segments with stable UUID, start/end, source track, speaker ID, text, revision, `provisional/final`, overlap state, and completeness state.
+- [ ] **[MUST · FOUNDATION] MODEL-007** Preserve simultaneous speech instead of forcing overlapping sources into a false single-speaker order.
+- [ ] **[MUST · FOUNDATION] MODEL-008** Keep identity candidates and confirmations separate from the final display name.
+
+### 10.4 Person profiles
+
+- [ ] **[MUST · M3] MODEL-009** Define stable `PersonID`, editable display name, multiple confirmed embeddings/prototypes, source quality, microphone/room context, model fingerprint, created/updated dates, and consent state.
+- [ ] **[MUST · M3] MODEL-010** Invalidate or migrate embeddings when the speaker-embedding model/version changes.
+
+---
+
+## 11. Audio capture design
+
+### 11.1 Online call mode
+
+- [ ] **[MUST · V1] CAP-001** Capture selected meeting-application audio through ScreenCaptureKit `.audio` output.
+- [ ] **[MUST · V1] CAP-002** Capture the selected microphone as a separate `.microphone` output on the macOS 15 baseline.
+- [ ] **[MUST · V1] CAP-003** Resolve the microphone using the identifier type ScreenCaptureKit expects; do not assume the existing Core Audio UID is directly interchangeable with `AVCaptureDevice.uniqueID`.
+- [ ] **[MUST · V1] CAP-004** Use media presentation timestamps as the synchronization source, not callback arrival time or wall-clock `Date`.
+- [ ] **[MUST · V1] CAP-005** Exclude FluidVoice audio where supported so cues/notifications are not written into meeting audio.
+- [ ] **[MUST · V1] CAP-006** Preserve remote/system and microphone tracks separately through processing and audio export.
+- [ ] **[MUST · FOUNDATION] CAP-024** Decide whether V1 source selection uses Apple's `SCContentSharingPicker` or a custom application picker, and document the privacy, control, and maintenance reason for that choice.
+- [ ] **[MUST · V1] CAP-025** Consume `.audio` and `.microphone` sample buffers into separate writers; do not use a convenience recording output that mixes away track identity on the macOS 15 baseline.
+
+### 11.2 In-room mode
+
+- [ ] **[MUST · V1] CAP-007** Record the selected microphone through a mic-only Core Audio/AVFoundation backend sharing the durable writer, without requiring Screen/System Audio permission solely for in-room mode.
+- [ ] **[MUST · V1] CAP-008** Diarize all detected room speakers, including the local user.
+- [ ] **[MUST · V1] CAP-009** Never label the whole in-room microphone as **You**.
+
+### 11.3 Durability and resource behavior
+
+- [ ] **[MUST · V1] CAP-010** Write rolling finalized chunks plus an atomically replaced manifest under Application Support.
+- [ ] **[MUST · V1] CAP-011** Bound memory regardless of meeting duration; sample buffers must not accumulate in RAM.
+- [ ] **[MUST · V1] CAP-012** Keep capture callbacks minimal and file writes on dedicated serial queues.
+- [ ] **[MUST · V1] CAP-013** Use compressed archival tracks or another measured format that avoids multi-gigabyte-per-hour PCM while preserving ASR quality.
+- [ ] **[MUST · V1] CAP-014** Normalize/downsample for ASR during offline processing, not on the critical capture callback.
+- [ ] **[MUST · V1] CAP-015** Preflight available space and continue monitoring storage while recording.
+- [ ] **[MUST · V1] CAP-016** Record explicit chunk boundaries for device changes, sleep/wake, source-app reconnection, and format changes.
+- [ ] **[MUST · V1] CAP-017** Losing one source must not silently terminate the healthy source.
+- [ ] **[MUST · V1] CAP-018** A crash may lose the unfinalized active chunk, but must preserve all finalized chunks.
+
+### 11.4 Headphones, Bluetooth, and leakage
+
+- [ ] **[MUST · V1] CAP-019** Recommend headphones for online meetings because system audio remains digital and remote speech is less likely to leak into the microphone.
+- [ ] **[MUST · V1] CAP-020** Allow AirPods/Bluetooth microphones but warn that using a Bluetooth headset microphone can switch macOS to a lower-quality two-way mode.
+- [ ] **[MUST · V1] CAP-021** Recommend built-in/USB microphone plus Bluetooth headphones output when appropriate, without overriding user choice.
+- [ ] **[MUST · V1] CAP-022** Detect/report probable duplicate remote speech across the microphone and system tracks; do not claim automatic echo cancellation.
+- [ ] **[OPTIONAL · V1] CAP-023** Add timestamp/acoustic duplicate suppression only after evaluation proves it does not remove real overlapping speech.
+- [ ] **[MUST · V1] CAP-026** Ask whether the selected online-call microphone is personal or shared. Only a confirmed personal mic may support the default **You** attribution.
+
+---
+
+## 12. Offline English transcription pipeline
+
+- [ ] **[MUST · V1] PIPE-001** Process the dual-track session manifest rather than pretending a captured meeting is one imported file.
+- [ ] **[MUST · V1] PIPE-002** Remote/application track: diarize, preserve cluster IDs/embeddings/quality, and transcribe bounded turns.
+- [ ] **[MUST · V1] PIPE-003** Microphone track: detect speech turns and correlate them against remote/system speech. Label clean personal-mic speech as **You**; mark probable leakage, shared-mic speech, or ambiguous overlap as **Microphone / Unknown** rather than misattributing it.
+- [ ] **[MUST · V1] PIPE-004** Merge tracks chronologically from their shared timestamp mapping while preserving overlaps.
+- [ ] **[MUST · V1] PIPE-005** Preserve raw meeting speech; do not automatically apply dictation filler removal, spoken-punctuation rewriting, or AI enhancement.
+- [ ] **[MUST · V1] PIPE-006** Reuse the selected safe file-transcription provider/model path to avoid duplicate model memory, with explicit serialization against dictation.
+- [ ] **[MUST · V1] PIPE-007** Checkpoint processing at track/chunk/phase boundaries so relaunch can resume.
+- [ ] **[MUST · V1] PIPE-008** Batch turns carefully; avoid one expensive ASR startup per tiny diarization segment.
+- [ ] **[MUST · V1] PIPE-009** Treat no-speech, incomplete-track, diarization failure, and ASR failure as distinct outcomes.
+- [ ] **[MUST · V1] PIPE-010** Fall back to a complete unlabeled transcript when speaker attribution cannot be completed safely.
+- [ ] **[MUST · V1] PIPE-011** On Intel, produce plain remote transcript plus track-origin distinction where possible; do not invent remote speaker labels.
+- [ ] **[MUST · FOUNDATION] PIPE-012** Every ASR request reads `languageCode` from its session; adding a future transcription-language picker requires no session-schema migration. Product UI localization remains separate.
+- [ ] **[MUST · V1] PIPE-013** Stitch remote speaker clusters across writer rotation, sleep/reconnect, device/format boundaries, and processing chunks using session-wide embeddings so the same person retains one speaker ID.
+- [ ] **[MUST · V1] PIPE-014** Pin ASR provider/model/version, diarization model fingerprint, normalization settings, and pipeline version to each processing attempt/checkpoint.
+- [ ] **[MUST · V1] PIPE-015** Dictation may preempt/pause meeting processing at a safe checkpoint; prove no concurrent provider mutation, model unload, or partial-result corruption.
+
+---
+
+## 13. Speaker identity and correction
+
+### 13.1 V1 session identity
+
+- [ ] **[MUST · V1] ID-001** Clean speech from a confirmed personal online-call microphone is labeled the configured local name plus **(You)**, or simply **You**; shared/ambiguous/leaked speech is not.
+- [ ] **[MUST · V1] ID-002** Remote speakers begin as stable generic labels ordered by first appearance.
+- [ ] **[MUST · V1] ID-003** Renaming a speaker changes display metadata, not segment IDs or cluster identity.
+- [ ] **[MUST · V1] ID-004** Global rename, segment reassignment, speaker merge, and undo remain available without voice profiles.
+
+### 13.2 Compact candidate confirmation
+
+- [ ] **[MUST · M3] ID-005** Show one unresolved confirmation per unique session speaker, not beside every passage.
+- [ ] **[MUST · M3] ID-006** **Yes** renames all linked segments and records an explicit confirmation event.
+- [ ] **[MUST · M3] ID-007** **No** rejects only that candidate for that session and does not permanently poison either profile.
+- [ ] **[MUST · M3] ID-008** No answer leaves the generic speaker label intact.
+- [ ] **[MUST · M3] ID-009** Use wording such as “Was this Erik?” rather than uncalibrated percentages.
+
+### 13.3 Cross-meeting voice profiles
+
+- [ ] **[MUST · M3] ID-010** Persistent profiles are explicit opt-in before the first embedding is stored.
+- [ ] **[MUST · M3] ID-011** Learn only from user-confirmed, sufficiently long, clean, non-overlapping speech.
+- [ ] **[MUST · M3] ID-012** Store several representative samples/prototypes per person across microphones and rooms.
+- [ ] **[MUST · M3] ID-013** Suggest a match only when it clears an evaluated absolute threshold and a margin over the second-best person.
+- [ ] **[MUST · M3] ID-014** Preserve an unknown-person outcome; never force the nearest profile.
+- [ ] **[MUST · M3] ID-015** Require confirmation again in each meeting until reliability evidence supports a different policy.
+- [ ] **[MUST · M3] ID-016** Support same-name people through distinct stable person IDs.
+- [ ] **[MUST · M3] ID-017** Provide rename, unlink, forget-one-person, and delete-all-profile controls.
+- [ ] **[OPTIONAL · M3] ID-018** Remember aliases by recurring meeting series after a separate product decision.
+
+### 13.4 Visual active-speaker assistance
+
+- [ ] **[OPTIONAL · M4] ID-019** Use platform UI as a candidate-name signal, never as the basis of diarization.
+- [ ] **[CONDITIONAL MUST · M4] ID-020** Prefer accessibility-exposed participant labels before OCR where reliable.
+- [ ] **[CONDITIONAL MUST · M4] ID-021** Sample low-frequency frames only around remote speech activity and crop to likely participant regions.
+- [ ] **[CONDITIONAL MUST · M4] ID-022** Keep frames in memory only; persist only candidate name, time interval, confidence, and platform-rule version.
+- [ ] **[CONDITIONAL MUST · M4] ID-023** Require agreement across multiple frames and align the visual interval with the diarized cluster.
+- [ ] **[CONDITIONAL MUST · M4] ID-024** Suppress suggestions during overlap, screen sharing, minimized/hidden UI, ambiguous highlights, or voice/visual disagreement.
+- [ ] **[CONDITIONAL MUST · M4] ID-025** Fall back silently to voice/manual identity when visual extraction is unavailable.
+- [ ] **[CONDITIONAL MUST · M4] ID-026** Obtain separate opt-in before examining the selected meeting window; restrict analysis to that window and likely participant/name regions, and skip shared-screen content.
+
+---
+
+## 14. Supported meeting platforms
+
+### 14.1 Capability model
+
+- [ ] **[MUST · V1] PLATFORM-001** Track support separately for **Recording**, **Meeting Detection**, **Visual Name Suggestions**, and **Voice Recognition**.
+- [ ] **[MUST · V1] PLATFORM-002** Define tiers:
+ - **Optimized:** claimed capabilities validated against the release matrix.
+ - **Compatible:** reliable manual recording/transcription with limited automation.
+ - **Beta:** expected to work but not sufficiently validated.
+ - **Unsupported:** known capture failure.
+- [ ] **[MUST · V1] PLATFORM-003** A platform UI change may downgrade name suggestions without disabling recording.
+- [ ] **[MUST · V1] PLATFORM-004** Expose exact tested native app/browser variants rather than one broad platform claim.
+
+### 14.2 Planned initial matrix
+
+| Target | Initial planned claim | Capture | Detection | Name suggestions | Required caveat |
+|---|---|---:|---:|---:|---|
+| Zoom Workplace desktop | Planned target | Planned—validation required | Planned | Planned | Validate helper/subprocess capture and UI variants |
+| Google Meet in Chrome | Planned target | Planned—validation required | Planned | Planned | Same-browser audio may be included |
+| Microsoft Teams desktop | Planned target | Planned—validation required | Planned | Planned | Validate current Teams helper/subprocess behavior |
+| Google Meet in Safari/Edge/Firefox | Planned compatible target | Planned—validation required | Best effort | Planned voice/manual | Treat each browser as its own variant |
+| Webex desktop | Planned compatible target | Planned—validation required | Best effort | Planned voice/manual | UI adapter must pass separately |
+| Slack desktop Huddles | Planned compatible target | Planned—validation required | Best effort | Planned voice/manual | Browser Huddles are separate variants |
+| FaceTime | Planned beta target | Planned—validation required | Best effort | Planned voice/manual | Validate capture/protected-content behavior first |
+| Other app/browser | Manual fallback target | Unknown until selected and checked | No promise | Planned voice/manual | Never imply compatibility before source health is proven |
+
+- [ ] **[MUST · V1] PLATFORM-005** Provide **Choose application…** for unrecognized meeting software.
+- [ ] **[MUST · V1] PLATFORM-006** Warn that ScreenCaptureKit exposes no browser-tab audio selector, so unrelated audio from the same browser may be included; do not claim identical behavior across all browsers without validation.
+- [ ] **[MUST · V1] PLATFORM-007** Validate Zoom/Teams helper processes; bundle-name matching alone must not earn an optimized claim.
+- [ ] **[MUST · V1] PLATFORM-008** Keep platform capture configuration separate from ASR/diarization configuration unless measurements prove an app-specific transcription benefit.
+- [ ] **[MUST · M4] PLATFORM-009** Version platform profiles by native/browser variant and known UI generation.
+- [ ] **[OPTIONAL · M4] PLATFORM-010** Store last-validated version/date in diagnostics without reporting meeting contents.
+- [ ] **[MUST · V1] PLATFORM-011** Show **Works best with** only for validated recording targets, followed by detailed per-capability status. Never present a planned capability as currently supported.
+
+---
+
+## 15. Meeting detection and suggestion prompt
+
+- [ ] **[OPTIONAL · V1] DETECT-001** Ship detection only after manual capture is trustworthy; default milestone is M4.
+- [ ] **[MUST · M4] DETECT-002** Make detection opt-in globally and suppressible per application.
+- [ ] **[MUST · M4] DETECT-003** Require sustained multi-signal evidence, never a window title alone.
+ - Candidate signals: known app/window, process audio input/output activity, active duration, and platform-specific context.
+- [ ] **[MUST · M4] DETECT-004** Debounce startup and show at most one prompt per meeting lifecycle.
+- [ ] **[MUST · M4] DETECT-005** **Not now** suppresses until that meeting ends.
+- [ ] **[MUST · M4] DETECT-006** Offer **Don't suggest for this app**.
+- [ ] **[MUST · M4] DETECT-007** Use **Set up** instead of **Record** when permission or source configuration is incomplete.
+- [ ] **[MUST · M4] DETECT-008** Merely presenting the suggestion never activates FluidVoice or steals keyboard focus. An explicit **Set up** action may activate the Meeting tab.
+- [ ] **[MUST · M4] DETECT-009** Notch Macs use a compact notch expansion.
+- [ ] **[MUST · M4] DETECT-010** Non-notch Macs use a top-center HUD on the meeting window's display, not literal screen center or pointer location.
+- [ ] **[MUST · M4] DETECT-011** Prompt timeout pauses for hover, keyboard focus, and VoiceOver interaction.
+- [ ] **[MUST · M4] DETECT-012** Meeting prompts yield to active dictation/command overlays and never delay dictation's first PCM.
+- [ ] **[OPTIONAL · M4] DETECT-013** Leave a subtle Meeting sidebar badge after timeout.
+- [ ] **[MUST · M4] DETECT-014** Use concise platform-aware copy:
+ - Normal: **“Looks like you're in a Zoom meeting. Record locally?”** with **Record** and **Not now**.
+ - First use: the same context with **Set up** and **Not now**.
+ - Include the detected app icon without implying FluidVoice has already begun capture.
+- [ ] **[MUST · M4] DETECT-015** Provide a keyboard-complete alternative through the FluidVoice menu-bar menu or Meeting tab. The nonactivating HUD may expose pointer and VoiceOver actions without pretending it accepts normal keyboard focus.
+
+---
+
+## 16. Menu-bar lifecycle and notifications
+
+### 16.1 Recording
+
+```text
+● Stop Meeting Recording
+ Recording Zoom · 24:18
+ Open Meeting Transcription
+────────────────────────────
+ Existing FluidVoice menu…
+```
+
+- [ ] **[MUST · V1] MENU-001** Extend the existing FluidVoice status item; do not create a second menu-bar icon.
+- [ ] **[MUST · V1] MENU-002** Show the FV icon with a recording mark that does not rely on red alone.
+- [ ] **[MUST · V1] MENU-003** Make **Stop Meeting Recording** the first enabled menu item while capture is active.
+- [ ] **[MUST · V1] MENU-004** Show platform/source, elapsed time, and **Open Meeting Transcription** next.
+- [ ] **[MUST · V1] MENU-005** Compute elapsed time when the menu opens; do not run a permanent one-second UI timer.
+- [ ] **[MUST · V1] MENU-006** Stop ends recording immediately and begins processing.
+
+### 16.2 Processing, interruption, and completion
+
+- [ ] **[MUST · V1] MENU-007** Replace the red recording state immediately after audio capture stops.
+- [ ] **[MUST · V1] MENU-008** Use a distinct amber/processing mark and first row **Transcribing meeting…**.
+- [ ] **[MUST · V1] MENU-009** Interrupted capture changes the first action to **Recording interrupted — Open Meeting**.
+- [ ] **[MUST · V1] MENU-010** Completion restores the normal icon and posts **Meeting transcript ready**.
+- [ ] **[MUST · V1] MENU-011** Temporarily expose **Open Latest Meeting Transcript** first after completion.
+- [ ] **[MUST · V1] MENU-012** Provide meaningful accessibility values for Recording, Processing, Interrupted, and Ready.
+- [ ] **[MUST · V1] MENU-013** Warn on Quit while recording with two clear choices: cancel Quit and keep recording, or **Stop & Quit — Transcribe Next Launch**, which persists `processingPending` for relaunch.
+- [ ] **[OPTIONAL · V1] MENU-014** Add a global Stop Meeting Recording shortcut after conflict review.
+
+---
+
+## 17. Permissions, consent, privacy, and retention
+
+- [ ] **[MUST · V1] PRIV-001** Track microphone and Screen/System Audio permission independently: unknown, granted, denied, restricted, and action required.
+- [ ] **[MUST · V1] PRIV-002** Explain why each permission is needed before presenting the system flow.
+- [ ] **[MUST · V1] PRIV-003** Keep a visible FluidVoice recording indicator for the complete capture duration; it supplements rather than replaces macOS privacy indicators.
+- [ ] **[MUST · V1] PRIV-004** Offer concise pre-start copy reminding the user to ensure participants know the meeting is being recorded.
+- [ ] **[MUST · V1] PRIV-005** Do not claim FluidVoice determines recording legality.
+- [ ] **[MUST · V1] PRIV-006** Exclude audio, transcript text, participant names, meeting titles, and embeddings from analytics/logs/uploads.
+- [ ] **[MUST · V1] PRIV-007** Define retention before release: delete after transcription, 7 days, 30 days, or never.
+- [ ] **[MUST · V1] PRIV-008** Distinguish **Delete Audio**, **Delete Meeting**, **Forget Person**, and **Delete All Meeting Data**.
+- [ ] **[MUST · V1] PRIV-009** Keep transcript usable after audio-only deletion.
+- [ ] **[MUST · V1] PRIV-010** Store sensitive files with owner-only permissions and accurate local-storage copy.
+- [ ] **[MUST · V1] PRIV-011** Do not claim app-managed encryption unless authenticated encryption and Keychain-backed key management are actually implemented.
+- [ ] **[OPTIONAL · SECURITY] PRIV-012** Add app-managed at-rest encryption after defining streaming/recovery behavior and migration.
+- [ ] **[MUST · M3] PRIV-013** Voice-profile opt-in and deletion are independent of meeting retention.
+- [ ] **[MUST · V1] PRIV-014** Warn that exported text, JSON, or audio leaves FluidVoice's storage controls and follows the destination's permissions/sync behavior.
+- [ ] **[MUST · V1] PRIV-015** Deletion covers active/partial chunks, finalized tracks, manifests, checkpoints, visual candidates, and recovery remnants in the selected scope; do not promise secure erase on APFS.
+- [ ] **[MUST · M3] PRIV-016** Complete a separate biometric/privacy review before remembered third-party voice profiles ship; require explicit opt-in, local-only storage, no sync, and dedicated disclosure.
+- [ ] **[MUST · M3] PRIV-017** Protect persisted voice profiles with authenticated encryption using a Keychain-held key, or explicitly re-review and approve the weaker storage risk before M3.
+
+---
+
+## 18. Reliability, recovery, and conflicts
+
+- [ ] **[MUST · V1] REL-001** On launch, scan sessions left in recording/stopping/processing states.
+- [ ] **[MUST · V1] REL-002** Salvage finalized chunks, mark unexpected capture termination as interrupted, and never auto-delete evidence.
+- [ ] **[MUST · V1] REL-003** Offer **Resume Processing**, **Reveal/Export Audio**, and **Delete** for recoverable sessions.
+- [ ] **[MUST · V1] REL-004** Handle force quit, power loss, sleep/wake, selected-app exit/relaunch, permission revocation, and disk exhaustion.
+- [ ] **[MUST · V1] REL-005** Handle microphone unplug/reconnect, Bluetooth route changes, and sample-rate changes without corrupting prior chunks.
+- [ ] **[MUST · V1] REL-006** Handle waiting room → meeting, breakout room/window, browser refresh, and source subprocess changes without silent loss.
+- [ ] **[MUST · V1] REL-007** Surface independent last-sample time, level, discontinuity, and failure for each track.
+- [ ] **[MUST · V1] REL-008** Meeting recording receives an explicit audio-activity lease.
+- [ ] **[MUST · V1] REL-009** V1 dictation hotkeys during meeting recording show a clear “Stop meeting recording first” message rather than opening a competing microphone session.
+- [ ] **[MUST · V1] REL-010** Meeting processing serializes safe access to shared ASR models and never corrupts dictation state.
+- [ ] **[MUST · V1] REL-011** No meeting prompt, menu update, or indicator adds latency to normal dictation capture when meeting recording is inactive.
+- [ ] **[MUST · V1] REL-012** Start, Stop, menu Stop, Quit, and late ScreenCaptureKit/writer callbacks are idempotent and generation-checked; repeated Stop enters disabled **Stopping…** and creates exactly one finalization/processing job.
+- [ ] **[MUST · V1] REL-013** Critical-low-disk behavior safely stops and finalizes capture before exhausting the reserve, then notifies the user without deleting audio.
+- [ ] **[MUST · V1] REL-014** Processing settings/provider changes cannot mutate an in-flight attempt; a deliberate retry creates a new pinned attempt while preserving prior checkpoints/results.
+
+---
+
+## 19. Performance, accessibility, and macOS behavior
+
+### 19.1 Performance
+
+- [ ] **[MUST · V1] PERF-001** Capture CPU, memory, and disk queues remain bounded for a one-hour session.
+- [ ] **[MUST · V1] PERF-002** Offline models remain unloaded during capture unless already needed elsewhere.
+- [ ] **[MUST · V1] PERF-003** Visual-name assistance, when enabled, runs at a bounded low rate and never competes with audio writes.
+- [ ] **[MUST · V1] PERF-004** Transcript rendering remains responsive with hours of segments; do not render an unbounded monolithic SwiftUI tree.
+- [ ] **[OPTIONAL · M4] PERF-005** Reduce or pause visual sampling under thermal or Low Power Mode pressure.
+
+### 19.2 Accessibility and Mac conventions
+
+- [ ] **[MUST · V1] A11Y-001** Every primary action is keyboard reachable.
+- [ ] **[MUST · V1] A11Y-002** Speaker identity uses explicit **Yes/No** text buttons, not icon-only controls.
+- [ ] **[MUST · V1] A11Y-003** VoiceOver has a logical focus order and concise recording/processing/error announcements.
+- [ ] **[MUST · V1] A11Y-004** Recording, processing, and failures never rely on color alone.
+- [ ] **[MUST · V1] A11Y-005** Respect Reduce Motion, Increase Contrast, and Reduce Transparency.
+- [ ] **[MUST · V1] A11Y-006** Transcript text supports standard selection, copy, find, context menus, and undo behavior.
+- [ ] **[MUST · V1] A11Y-007** Prompt/HUD behavior works across full screen, Spaces, and multiple displays without focus theft.
+- [ ] **[MUST · V1] A11Y-008** The menu-bar workflow remains complete when the main window is closed.
+- [ ] **[MUST · V1] A11Y-009** Reuse existing FluidVoice theme typography, cards, buttons, spacing, and control heights; avoid one-off chrome.
+
+---
+
+## 20. Future live-mode seam
+
+- [ ] **[MUST · FOUNDATION] LIVE-001** Coordinator supports transcript mode `.offlineAfterStop` now and `.live` later.
+- [ ] **[MUST · FOUNDATION] LIVE-002** Transcript canvas accepts append, revise, finalize, merge, reassignment, and identity-change events for stable IDs.
+- [ ] **[MUST · FOUNDATION] LIVE-003** Capture and processing communicate through bounded streams with backpressure; live failure cannot stop durable recording.
+- [ ] **[FUTURE · LIVE] LIVE-004** Feed downsampled chunks to a live ASR pipeline while archival writes remain independent.
+- [ ] **[FUTURE · LIVE] LIVE-005** Show provisional text/speaker state without excessive visual noise.
+- [ ] **[FUTURE · LIVE] LIVE-006** Auto-scroll only while the user is already at the bottom.
+- [ ] **[FUTURE · LIVE] LIVE-007** Revisions preserve reading position, selection, and manual corrections.
+- [ ] **[FUTURE · LIVE] LIVE-008** Stop runs final offline reconciliation and atomically finalizes provisional segments.
+- [ ] **[FUTURE · LIVE] LIVE-009** Apply thermal/memory backpressure and disable live mode clearly on unsupported hardware.
+
+---
+
+## 21. Acceptance and release gates
+
+### 21.1 Platform matrix
+
+- [ ] **[MUST · V1] TEST-PLATFORM-001** For every claimed platform/variant: waiting room, join/rejoin, mute, meeting end, app quit/relaunch, screen share, window resize, minimized/occluded window, breakout room, and app update.
+- [ ] **[MUST · V1] TEST-PLATFORM-002** Browser test documents unrelated-tab audio behavior and verifies the warning is accurate.
+- [ ] **[MUST · V1] TEST-PLATFORM-003** Zoom/Teams tests prove actual meeting/helper audio reaches the saved system track.
+- [ ] **[MUST · V1] TEST-PLATFORM-004** No individual capability receives **Optimized** until that capability's own matrix passes; unavailable detection/name features do not block an independently validated Recording status.
+
+### 21.2 Audio matrix
+
+- [ ] **[MUST · V1] TEST-AUDIO-001** Built-in mic, USB mic, wired headphones, AirPods output, AirPods microphone, device unplug/reconnect, and route change.
+- [ ] **[MUST · V1] TEST-AUDIO-002** Headphones and laptop speakers with at least two remote speakers, overlap, silence, and notification/music interference.
+- [ ] **[MUST · V1] TEST-AUDIO-003** A long-session test proves both tracks remain within the numeric drift and unexplained-gap budgets set by TEST-BUDGET-001/002.
+- [ ] **[MUST · V1] TEST-AUDIO-004** Source meter/health UI agrees with the audio actually written to disk.
+
+### 21.3 Failure and recovery matrix
+
+- [ ] **[MUST · V1] TEST-FAIL-001** Force quit, crash-equivalent termination, sleep/wake, source app quit/relaunch, microphone removal, capture permission revocation, and disk exhaustion.
+- [ ] **[MUST · V1] TEST-FAIL-002** Every finalized chunk before failure is recoverable.
+- [ ] **[MUST · V1] TEST-FAIL-003** Processing failure preserves audio and retry resumes from a durable checkpoint.
+- [ ] **[MUST · V1] TEST-FAIL-004** Window close, tab navigation, and menu-bar-only use do not stop recording.
+- [ ] **[MUST · V1] TEST-FAIL-005** Permission lifecycle covers first grant, denial, return from System Settings, revocation, regrant, and relaunch for microphone and Screen/System Audio independently.
+- [ ] **[MUST · V1] TEST-FAIL-006** Duplicate Start/Stop/menu/Quit actions and stale callbacks produce exactly one session, one finalized manifest, and one queued processing job.
+
+### 21.4 Speaker and identity matrix
+
+- [ ] **[MUST · V1] TEST-ID-001** Remote speaker labels remain stable through rename, merge, export, persistence, and reopen.
+- [ ] **[MUST · M3] TEST-ID-002** Establish a minimum precision target before beta; optimize for fewer trustworthy suggestions rather than broad incorrect coverage.
+- [ ] **[MUST · M3] TEST-ID-003** Test unknown people, same-name people, different microphones/rooms, overlapping speech, and visually ambiguous names.
+- [ ] **[MUST · M3] TEST-ID-004** **No** does not alter stored profiles; only explicit **Yes** may enroll clean samples.
+- [ ] **[MUST · M3] TEST-ID-005** Forget Person removes the profile without corrupting historical transcript display.
+- [ ] **[MUST · V1] TEST-ID-006** A remote speaker appearing across multiple processing chunks retains one session-scoped speaker identity unless the evidence is genuinely ambiguous.
+
+### 21.5 Privacy and accessibility matrix
+
+- [ ] **[MUST · V1] TEST-PRIV-001** Inspect storage, analytics, logs, diagnostics, and network traffic for audio/transcript/name leakage.
+- [ ] **[MUST · M4 if visual enabled] TEST-PRIV-002** Confirm no meeting screenshots/video frames persist.
+- [ ] **[MUST · V1] TEST-A11Y-001** Complete the core workflow with keyboard only and VoiceOver.
+- [ ] **[MUST · V1] TEST-A11Y-002** Verify high contrast, reduced motion, reduced transparency, multiple displays, Spaces, and fullscreen meetings.
+
+### 21.6 Repository and release proof
+
+- [ ] **[MUST · V1] TEST-REPO-001** Run SwiftFormat and strict SwiftLint before commit-ready feature work.
+- [ ] **[MUST · V1] TEST-REPO-002** Validate implementation builds with `sh build_with_FI_incremental.sh` and inspect the installed `/Applications/FluidVoice.app`.
+- [ ] **[MUST · V1] TEST-REPO-003** Do not treat Debug/XCTest-host runs as release proof; use focused logic checks plus installed-app, real-meeting evidence.
+- [ ] **[MUST · V1] TEST-REPO-004** Verify the minimum supported macOS path and low-resource Apple Silicon behavior; describe Intel as reduced capability until separately proven.
+- [ ] **[MUST · V1] TEST-REPO-005** Update the requested release-notes version locally, credit contributors as required, and never commit ignored release notes.
+- [ ] **[MUST · V1] TEST-REPO-006** Run real Intel capture plus plain English transcription before claiming Intel meeting support; universal build slices alone are insufficient.
+
+### 21.7 Measurable budgets to set before implementation
+
+- [ ] **[MUST · FOUNDATION] TEST-BUDGET-001** Set maximum supported meeting duration and the long-session validation duration; one hour alone is not the product limit unless explicitly chosen.
+- [ ] **[MUST · FOUNDATION] TEST-BUDGET-002** Set numeric limits for accumulated track drift, unexplained audio gap, source-loss alert latency, and maximum recoverable loss from the active chunk.
+- [ ] **[MUST · FOUNDATION] TEST-BUDGET-003** Set disk reserve and warning/safe-stop thresholds.
+- [ ] **[MUST · FOUNDATION] TEST-BUDGET-004** Set CPU, RAM, buffer depth, and disk-write budgets for recording on the lowest supported Apple Silicon class.
+- [ ] **[MUST · FOUNDATION] TEST-BUDGET-005** Set diarization quality and cross-chunk speaker-continuity targets before calling speaker separation trustworthy.
+- [ ] **[MUST · M3] TEST-BUDGET-006** Set identity-candidate precision, unknown rejection, and false-enrollment targets before M3 beta.
+- [ ] **[CONDITIONAL MUST · M4] TEST-BUDGET-007** Set visual sampling rate, CPU budget, candidate agreement window, and false-name target before visual assistance ships.
+
+---
+
+## 22. Open decisions for alignment
+
+These do not block writing the PRD. They must be resolved before their associated milestone begins.
+
+- [ ] **[MUST · FOUNDATION] OPEN-001 — Default audio retention.**
+ - Recommendation: 7 days, with Delete after transcription, 30 days, and Never.
+
+- [ ] **[MUST · FOUNDATION] OPEN-002 — Dictation during recording.**
+ - Recommendation for V1: keep meeting recording healthy and explain that dictation is unavailable until the meeting stops.
+
+- [ ] **[MUST · M3] OPEN-003 — Voice-profile opt-in scope.**
+ - Recommendation: explicit first-use explanation plus a per-person **Remember this voice** choice; never infer consent from rename alone.
+
+- [ ] **[MUST · M4] OPEN-004 — Detection release timing.**
+ - Recommendation: after trustworthy manual recording, not in the first vertical slice.
+
+- [ ] **[MUST · M4] OPEN-005 — Visual active-speaker release timing.**
+ - Recommendation: experimental after voice/manual identity works; platform UI analysis must never block capture.
+
+- [ ] **[MUST · V1] OPEN-006 — Exact optimized variants.**
+ - Recommendation: begin with Zoom desktop, Google Meet in Chrome, and current Teams desktop; add browsers/versions only after matrix proof.
+
+- [ ] **[MUST · V1] OPEN-007 — Intel product copy.**
+ - Recommendation: recording plus plain English transcript, with **You** versus **Meeting audio** when track origin permits, but no remote-speaker labels.
+
+- [ ] **[OPTIONAL · V1] OPEN-008 — Pause/resume.**
+ - Recommendation: defer unless user research makes it essential.
+
+- [ ] **[MUST · V1] OPEN-009 — Stop cue.**
+ - Recommendation: no audible cue routed into meeting/application audio; use visual state and optional haptic-equivalent UI only.
+
+- [ ] **[MUST · V1] OPEN-010 — Quit behavior.**
+ - Recommendation: closing the window continues; quitting prompts **Keep Recording** (cancel Quit), **Stop & Quit**, or Cancel.
+
+- [ ] **[MUST · V1] OPEN-011 — Local display name source.**
+ - Recommendation: a meeting-specific preferred name stored in FluidVoice, editable from the transcript participant chip.
+
+- [ ] **[MUST · V1] OPEN-012 — Backup policy.**
+ - Decide whether recordings, transcripts, and voice profiles participate in FluidVoice backup/export independently.
+
+---
+
+## 23. Highest-risk implementation traps
+
+- [ ] **[MUST · FOUNDATION] RISK-001** ScreenCaptureKit microphone identifiers may not equal the current Core Audio device UID.
+- [ ] **[MUST · FOUNDATION] RISK-002** The current diarization wrapper discards the embeddings needed for remembered identity.
+- [ ] **[MUST · FOUNDATION] RISK-003** Current speaker-segment IDs change when the speaker name changes.
+- [ ] **[MUST · FOUNDATION] RISK-004** Current file-history/UserDefaults storage is unsuitable for long recordings and crash recovery.
+- [ ] **[MUST · FOUNDATION] RISK-005** A whole-file local-mic transcript cannot be accurately merged with remote turns without speech timestamps.
+- [ ] **[MUST · FOUNDATION] RISK-006** Reusing `ASRService.start()` would entangle meeting capture with dictation streaming, media pausing, and teardown.
+- [ ] **[MUST · FOUNDATION] RISK-007** Existing menu-bar `isRecording` means dictation only, and the current template icon does not actually draw the requested red dot.
+- [ ] **[MUST · FOUNDATION] RISK-008** Changing a ScreenCaptureKit stream/device configuration mid-session can create interruption; represent it as a chunk boundary.
+- [ ] **[MUST · V1] RISK-009** Browser capture is application-wide, and Zoom/Teams helper-process behavior needs real proof.
+- [ ] **[MUST · V1] RISK-010** Meeting recovery cannot depend on the existing graceful shutdown timeout.
+- [ ] **[MUST · M4] RISK-011** Active-speaker visuals are brittle across layout changes, sharing, overlap, and minimized windows.
+- [ ] **[MUST · M3] RISK-012** Incorrect profile enrollment compounds future errors; only confirmed clean speech may update a profile.
+
+---
+
+## 24. Primary references
+
+- Apple ScreenCaptureKit: <https://developer.apple.com/documentation/screencapturekit>
+- Apple `SCContentFilter`: <https://developer.apple.com/documentation/screencapturekit/sccontentfilter>
+- Apple ScreenCaptureKit application-level audio discussion: <https://developer.apple.com/videos/play/wwdc2022/10155/>
+- Apple ScreenCaptureKit introduction: <https://developer.apple.com/videos/play/wwdc2022/10156/>
+- Apple Core Audio process input activity: <https://developer.apple.com/documentation/coreaudio/kaudioprocesspropertyisrunninginput>
+- Apple Bluetooth headset quality behavior: <https://support.apple.com/en-us/102217>
+- Google Meet requirements: <https://support.google.com/meet/answer/7317473>
+- Google Meet layouts: <https://support.google.com/meet/answer/9292748>
+- Zoom Workplace desktop app: <https://support.zoom.com/hc/en/article?id=zm_kb&sysparm_article=KB0064516>
+- Slack Huddles: <https://slack.com/help/articles/4402059015315-Use-huddles-in-Slack>
+- Webex system requirements: <https://help.webex.com/en-us/article/nk90t65/Webex-App-System-Requirements-and-Support-Policy>