With Rev AI streaming, send audio as binary messages and end with a single EOS text message; do not send other text. Render partial hypotheses as live updates and commit only final hypotheses, which carry the timestamps and confidence scores.

Context: Official Rev AI docs (streaming overview): documents the message protocol agents must implement. The client sends binary WebSocket messages with audio chunks; the only text message the client may send is EOS to gracefully end the stream. The server returns connected, partial, and final hypotheses. Partials are unstable by design; only final hypotheses carry per-element timestamps and confidence. Plan your UI updates around partials and your persistence around finals.