Training Data Contribution Risk

The risk is usually addressed contractually — a clause stating that submitted data will not be used for training — and contracts are the weakest control available. They change at renewal, they are interpreted by the party that wrote them, and compliance is unverifiable from outside. The structural alternative is not to submit the material. That reframes the question from what does the contract say to what actually leaves the environment, which is answerable by architecture rather than by assurance.

Where Centralpoint runs embedded local models — Llama, Qwen or ONNX inside the organization's own infrastructure — prompts and retrieved content are never transmitted, so the contribution question does not arise. Where cloud providers are used, the organization's prompts and skills still reside in its own environment rather than in the provider's tooling, so the encoded judgement stays local regardless of which model performs the inference.