AI Can Do Work. Humans Remain Responsible.

The models and knowledge behind the work.

What runs locally? What goes through OpenRouter? Are the documents we supply training data? Our AI models, tools and knowledge reference explains each part through the Laser TeamWork pilot, with dated labels for installed, tested and planned work.

Where assistance is useful.

AI can help organize source material, inspect code, propose an implementation, draft documentation and suggest tests. It is particularly useful when a person can define a bounded outcome and inspect the result.

We distinguish help during development from AI running inside a visitor-facing feature. The latter adds a service dependency with its own data, availability and failure questions.

What a person must decide.

• Which problem is worth solving and who it affects.
• Whether a factual claim has adequate evidence.
• What information may be shared with an external service.
• Whether code and behavior satisfy the requirement.
• Whether the change is authorized for production.

An assistant’s confidence is not a substitute for these decisions. A second generated opinion is not automatically an independent review.

What verification looks like.

For a product claim, consult the primary documentation or the actual source. For an interface, complete the important journey. For a change to code, inspect the diff and test the behavior it changes. For a deployment, verify the named destination.

Record the result in a form a successor can use. “Tested” needs to say what was tested and in which environment. “Reviewed” needs a real reviewer.

Production permission is explicit.

A successful local check does not authorize a live change. The person responsible for the system approves the release after reviewing its scope and destination. The deployment Lab explains why this boundary matters in our own work.

Keep durable knowledge available.

Requirements, important sources, business rules and source code should remain in ordinary project records. If an assistant becomes unavailable, the project should still explain itself.

Do not put credentials or unnecessary private records into prompts or published examples. Give the tool the minimum context needed for the task and use the approved handling process for the rest.

Tell readers what the evidence supports.

AI assistance is part of this project’s drafting and development workflow. That disclosure does not certify the result. Sources, test records and human approval provide separate evidence.

We correct unsupported claims of review or testing instead of treating a badge as harmless decoration. If you find a claim we cannot support, report it.