Catch the turn that broke.
Inspect interruptions, repeated actions, and the exact events behind each failure.
Findings
Deterministic checks · event evidence
Event timeline
NVIDIA model evidence
Recorded Nebius public-endpoint runs
Model output is validated separately. A rejected response is not a failed agent test. Read the recorded date and result for each sample; these are not live calls.
Measured audio replays
Whisper → offline reference agent → Kokoro reply · recorded test runs
Fault modes intentionally change the reference agent. File-based processing times include transcription and speech generation; they are not streaming interruption latency. No human participants.
Original audio samples
Locally generated Kokoro voices · synthetic, not human recordings
Change the date
“Wait, change the booking to Friday. Do not save Thursday.”
Cancel the booking
“Cancel that. Do not make any booking.”
File-based reference-agent replay is measured above. Live microphone streaming and human testing are not validated.