PASS Existing ranker works in the actual v1.2 document
PASS No model fetched before explicit consent
PASS Actual quantized Whisper detects the missing NOT on the synthetic speech fixture
PASS ASR comparison does not rewrite captions or approve cuts
PASS Speech record binds actual media bytes, transcript, model revision and range
PASS Changed captions invalidate an already completed speech check
PASS Explicit caption correction preserves timing and clears approvals
PASS Video-first transcription creates real estimated cues and applies only on request
PASS Generated captions feed the real selection and ZIP-export pipeline
PASS Actual browser speech recognition also processes a natural-speech fixture
PASS Cancel terminates speech work without applying a partial result
PASS Retrieval surfaces a related correction 180 seconds outside the cut without splicing it
PASS Evidence Desk fits a 390-pixel viewport
PASS No browser errors or audio-upload requests were observed
