Does the AI learn from what I type?

The contract does the work.

CrossCoach sends text to two providers, both through their commercial APIs: Anthropic, whose Claude models run the cross-examination and the evaluator, and OpenAI, whose embeddings turn short excerpts into vectors so we can find the relevant source passages. Commercial API traffic sits outside model training at both companies. Anthropic’s terms say it in one sentence:

“Anthropic may not train models on Customer Content from Services.”

— Anthropic Commercial Terms

“Customer Content” is defined in the same section as Inputs and Outputs: what you type, and what the model answers. OpenAI’s API terms make the equivalent commitment for API traffic.

Retention at the providers is brief and bounded.

The providers hold API inputs and outputs for a limited window, up to 30 days, for trust-and-safety and abuse monitoring, then delete them. At no point in that window is the material used for training. The model that cross-examines you next month is the same model everyone gets; nothing you typed has changed it.

What CrossCoach itself stores.

In standard practice mode, your transcript and evaluation are stored in our Sydney database (Supabase, ap-southeast-2, encrypted at rest), and the application runs in Sydney as well, so your data is stored and processed in Australia. That storage is for you. It is never used to train any model. Inside the product, your transcripts and scores are visible only to you; there is no manager or trainer view.

In private mode, CrossCoach stores nothing at all. The conversation lives in your browser’s memory; your feedback is generated and shown, never written down; close the tab and it is gone. The exchange still passes through the model under the same no-training terms; there is simply no CrossCoach record of it afterwards.