Build task-bound human review materialization

Add repeat playback and timed subspan approval

Constrain review controls to a compact panel

Move review controls below clip workspace

Apply autoplay to manual review navigation

Add configurable review auto-advance and autoplay

Build human review workstation with strict annotations

Implement source-accurate spans and Gemini audio verification

Operationalize calibration and checkpoint evaluation

Harden runtime portability and training safeguards

Add immutable stage resumes and evaluation workflows

Harden TTS data integrity and experiment lineage

Complete semantic and spelling gate baseline

Auto-load environment variables from .env file inside semantic gate validation

Implement LLM semantic meaningfulness gate option using Gemini 3.5 Flash with configurable thinking levels (MEDIUM default)

Implement spelling validator with Moratuwa verified wordlist lookup and recursive chunk trimming

Implement accuracy tracking and local text hydration in LLM CC pipeline

WIP: asr-audit faster-whisper and speaker promotion dashboard

Make inline ASR validation opt-in

Add promotion-aware dashboard metrics

Add SpeechBrain speaker embedding backend

Add subwindow speaker change scoring

Add optional speaker promotion stage

Document TTS and voice cloning roadmap

Keep exact caption boundaries as default

Add padded caption boundary mode for cleaner edges

Default to compact windowed chunking

Add speaker and music quality proxy gates

Add caption_raw boundary mode and make it default

Enforce non-overlapping refined clip timestamps

Add selective acquisition mode and refresh pipeline docs

Implement deterministic quality gates

Document design gaps implementation plan

Add ASR local hardening, rejected-eval extraction, and design gaps register

Add pilot_002 outputs and pending pipeline hardening changes

Enforce mandatory run dashboards and add master dashboard index

feat: complete LLM CC chunking pipeline with pilot results