What should an AI model maker show before calling a model 'aligned'?

OpenAI called GPT-6 Astra its ‘most aligned model yet’ when it announced the model last week. That phrase is doing more work than the launch. A buyer cannot use it to decide whether the model will stay inside a task boundary, flag uncertainty, or make a bad situation worse. I would rather see a small, ugly scorecard: cases it refused correctly, cases it over-refused, what changed after outside testing, and where the company still tells people not to use it. If ‘aligned’ cannot survive those questions, it is a compliment, not product information. What would you need before that label changed a real decision?