Responsible AI

Coach

Coach uses a language model to put your coaching into plain words. It does not use one to decide what is safe. Those are two different jobs, and Coach keeps them apart on purpose.

The model phrases, the engine decides

Every number that matters is computed by Coach's own rules, not by a model. Your readiness verdict, your heart-rate zones, the ceiling on a session and any clinical stop are worked out in code from your own data. The language model is only allowed to re-voice what the engine has already decided. It quotes the numbers; it never invents one, and it cannot change a verdict or lift a ceiling.

What is sent, and to whom

To write a reply, Coach sends the relevant context from your account, your recent numbers, your training, your goals and your message, to Anthropic, the maker of the Claude model, through our own server. Your device never holds a model key. Anthropic processes this to generate the reply and, under its commercial API terms, does not use it to train its models. We name Anthropic and our retention in the privacy policy.

Safety is enforced in code, not left to the model

A clinical warning and a heart-rate ceiling are hard rules. The model is checked against them and cannot soften, skip or override them. If Coach ever needs to tell you to stop or to see a professional, that comes from the rules, not from a model's judgement in the moment.

What the model can and cannot be relied on for

Questions about how Coach uses a model: hello@personal-coach.io