{
  "model_name": "gpt-realtime-2",
  "model_organization": "OpenAI",
  "submitting_organization": "Sierra",
  "submission_date": "2026-07-12",
  "submission_type": "custom",
  "modality": "voice",
  "contact_info": {
    "email": "soham@sierra.ai",
    "name": "Sierra Research Team"
  },
  "results": {
    "banking_knowledge": {
      "pass_1": 10.309278350515463,
      "retrieval_config": "alltools"
    }
  },
  "is_new": true,
  "trajectories_available": true,
  "trajectory_files": {
    "banking_knowledge": "gpt-realtime-2_voice_xhigh_usersim-gpt55xhigh_banking"
  },
  "methodology": {
    "evaluation_date": "2026-07-12",
    "tau2_bench_version": "1.0.1",
    "user_simulator": "gpt-5.5-2026-04-23 (xhigh)",
    "notes": "Custom submission: first voice-banking (voice + knowledge) entry. Identical to the standard voice evaluation pipeline (audio-native, regular speech complexity, ElevenLabs eleven_v3 TTS, agent reasoning_effort=xhigh) but run on the banking_knowledge domain with retrieval-config=alltools, and the user simulator LLM is gpt-5.5-2026-04-23 at reasoning_effort=xhigh instead of the leaderboard-standard v1.0 (gpt-4.1). Not comparable to standard-user-simulator leaderboard entries. Re-evaluated under tau2-bench v1.0.1 (banking_knowledge grading/task fixes #329/#388/#397/#402/#403): all 97 trajectories re-graded with the fixed evaluate-trajs path, and the 15 calls whose agents were exposed to the pre-fix environment (transaction ordering, Platinum Rewards doc title) re-run live under the fixed environment regardless of prior outcome. task_007 flipped to pass on re-run.",
    "verification": {
      "modified_prompts": false,
      "omitted_questions": false
    }
  },
  "voice_config": {
    "provider": "openai",
    "model": "gpt-realtime-2",
    "tick_duration_seconds": 0.2,
    "max_steps_seconds": 1200.0,
    "user_tts_provider": "elevenlabs/eleven_v3"
  },
  "model_release": {
    "release_date": "2026-05-07",
    "announcement_url": "https://openai.com/index/advancing-voice-intelligence-with-new-models-in-the-api/",
    "announcement_title": "Advancing voice intelligence with new models in the API"
  },
  "reasoning_effort": "xhigh",
  "interaction_metrics": {
    "version": "1.0",
    "config": {
      "tick_duration_sec": 0.2,
      "no_yield_window_sec": 2.0,
      "backchannel_yield_window_sec": 1.0,
      "vocal_tic_yield_window_sec": 1.0,
      "non_directed_yield_window_sec": 1.0,
      "vocal_tic_response_window_sec": 2.0,
      "non_directed_response_window_sec": 2.0
    },
    "domains": {
      "banking_knowledge": {
        "response_latency_mean": 2.036130136986302,
        "yield_latency_mean": 0.6110381077529566,
        "response_rate": 0.9442198868229588,
        "yield_rate": 0.9973787680209698,
        "agent_interruption_rate": 0.2667744543249798,
        "selectivity_backchannel": 0.03448275862068961,
        "selectivity_vocal_tic": 0.1544715447154471,
        "selectivity_non_directed": 0.1568627450980392,
        "counts": {
          "n_simulations": 97,
          "response_total": 1237,
          "yield_total": 763,
          "backchannel_total": 58,
          "vocal_tic_total": 123,
          "non_directed_total": 153,
          "agent_interrupts_count": 330
        }
      }
    },
    "overall": {
      "response_latency_mean": 2.036130136986302,
      "yield_latency_mean": 0.6110381077529566,
      "response_rate": 0.9442198868229588,
      "yield_rate": 0.9973787680209698,
      "agent_interruption_rate": 0.2667744543249798,
      "selectivity_backchannel": 0.03448275862068961,
      "selectivity_vocal_tic": 0.1544715447154471,
      "selectivity_non_directed": 0.1568627450980392,
      "counts": {
        "agent_interrupts_count": 330,
        "backchannel_total": 58,
        "n_simulations": 97,
        "non_directed_total": 153,
        "response_total": 1237,
        "vocal_tic_total": 123,
        "yield_total": 763
      }
    }
  }
}