| Google LLC — Vertex AI (Gemini) | Primary model for evaluating assessments: transcripts, recordings, submissions and drawings against the rubric | Assessment content, transcripts, recordings | United States (Vertex AI’s US multi-region endpoint) | No training on customer data (Google Cloud terms and Data Processing Addendum) |
| OpenAI, L.L.C. | Real-time voice conversations, text-to-speech, speech-to-text fallback, file analysis, optional multi-model grading | Student audio, transcripts, submitted files | United States | API data is not used for training; may be retained up to 30 days for abuse monitoring |
| ElevenLabs, Inc. | Conversational voice agents, speech-to-text, text-to-speech | Student audio, transcripts, agent responses | United States | Not used for training under our terms; request logs are currently retained — zero-retention mode being enabled |
| Anthropic, PBC | Optional multi-model grading (“council”), only when enabled | Transcripts, submission content | United States | API data is not used for training; limited retention for trust and safety |
| Deepgram, Inc. (voice service pilot) | Speech-to-text for Prova’s self-hosted voice service | Student audio | United States | Not used for training: every request opts out of Deepgram’s Model Improvement Program |
| OpenRouter, Inc. (voice service pilot) | Routes language-model requests to inference hosts in the US (e.g., DeepInfra, Baseten) | Conversation text, assessment context | United States (OpenRouter’s US region, which routes only to US inference endpoints) | Requests sent with data collection denied; no training |
