Code owns facts, the model owns voice
The one rule the whole system runs on, on a single screen.
A turn splits in two. Code owns the facts, what was actually generated, what a setting is set to, what a search really returned: and produces them rather than letting the model guess at them. The model owns the voice, the phrasing, the warmth, the judgment, which is the part it’s genuinely good at. Ask the model to own a fact and you’ve invited a confident, fluent lie. Keep the two on their own sides of the line and “it’s done” can’t be false, because the code that says so is the same code that did it.
How I learned this the hard way: an AI that invented its own memories four days into its life, is The wound that shaped everything. This is just the rule it left behind.