Agree the limits before the model is built.

Before any of your material is used for training, search or testing, we sort it with you.

• Sources that may be used
• Material that is excluded or restricted
• What happens to copies and anything derived from them
• What is returned or deleted when the work ends

We describe the model by its purpose, its limits and its weak spots.

• The tasks it is meant to do
• Outputs it must not produce
• Uses it is not meant for
• Cases that are sensitive or ambiguous

You check the work at four points before it reaches your people.

• The dataset, before training starts
• The evaluation, before deployment
• Sample outputs, before your staff or customers use them
• Switching on a new model or adapter

We do the engineering. You and your advisers make the business, domain and policy decisions.

• Engineering support from Random Walk
• Decisions about use from you
• Advisers where the matter calls for them
• Exceptions written down before work continues

Weak cases and open questions stay next to the work, where the next person to touch it will see them.

• Evaluation examples
• Notes on known limits
• Examples where the model fails
• Open questions and next steps

It describes how we work. It is not legal advice, an approval, a universal AI policy or a promise that every risk is gone.