Thanks, Aldo. This is a better method than the one in the post, and we agree with all of it. The 16-to-12 example was there to show the arithmetic of the 10-point rule, not to prove a loss, and one generation per cell is too noisy to call it on its own. Repeating each fixed prompt a few times in fresh sessions, and logging search mode, locale and model, is what turns a wobble into a signal. Keeping "recommended by name", "linked as a source" and "described accurately" as separate columns is a great call too, because they move independently. And you're right that the answer to "why didn't you include us?" is a generated explanation, not a log of the model's reasoning. That's why the post says to test each reason against your own site before acting on it.
