Sunny — Adaptive Real-Time Voice Companion

AssemblyAI supplies real-time SpeechStarted and final-turn events. Sunny handles the accepted turn, controls response playback, and yields when new human speech arrives. In a physical-device session, a human microphone passed through AssemblyAI and Sunny to speaker output. A separate session recorded true human barge-in: SpeechStarted triggered playback stop, followed by a final turn and new response. The redacted completion record and event ordering are public; equal event timestamps do not establish zero acoustic latency. An existing offline RhythmAdapter is a separate evidence track. Its source-backed numerical replay uses synthetic 900 ms pause observations to update a bounded next-response pause policy from 300 to 360 to 414 ms. These are computed policy values, not measured performance or proof of live personalization. The offline module is not verified as integrated with the human microphone path. Sunny does not claim identity, personality, voiceprint, or emotion inference. The public repository includes architecture, claim-by-claim evidence status, redacted human session records, an executable source excerpt, and contract tests. The 95-second submission film is a scripted product visualization with a labeled, sanitized real-microphone event-proof segment at 01:05–01:20. The animated woman is not the human tester; no tester audio or transcript is published. Baseline comparison, stop latency, and pause performance have not yet been benchmarked.