The two words “this time” in a dialog heading
Everything that took real time on the dual-AI role-assignment screen is invisible — one word, one button, one line of text, each pinned to a rule we will not break.
There are two words in the heading, and they are the only thing on this screen that really matters.
"This time"
We were building the dialog where two AIs are assigned their parts. The heading reads:
How should the two AIs divide the work this time?
Not "Configure AI roles." Not "Choose your development assistant." This time.
Those two words carry a rule we do not break:
- A role belongs to this round of work. It does not belong to the model.
GPT is not "the verifier." Claude is not "the builder." Gemini is not "the analyst." No model is permanently bound to an engineering role — the same model can be Builder in one task and Verifier in the next, but every change of role has to re-establish its own context, session and authority boundary.
If the heading said "Configure AI roles," the role would have become a property of the model. One word, and the architecture is already wrong.
The swap button in the middle
Between the two cards there is a "swap roles" control.
It is not there for convenience. It is the physical form of the rule above — at a glance you can see that nobody owns a role. You can exchange building and checking at any moment and the system is not confused, because it never remembered "which model does what" in the first place.
The line along the bottom
At the foot of the dialog:
The independent checker will not decide for you, and the builder cannot pass its own work. Every critical decision still has to be confirmed and submitted by you.
That sentence is placed there deliberately.
Not in the help documentation. Not deep inside a settings page. On the screen where you are about to press start.
Because it corresponds to our hardest rule:
- An AI saying "I am done" costs exactly as much as an AI saying "I was wrong." Which is nothing.
So completion is never declared by whoever did the work. A claim needs evidence, and the evidence has to be independently reproducible. The builder never gets acceptance authority; the checker never gets decision authority. Both sides can only narrow the question. The last move has to be made by a person.
What we deliberately did not build
Three things are missing from this screen on purpose:
- No "recommended pairing."
- No automatic assignment.
- No confidence score.
The system does not pick who builds. It can make the choice legible. It cannot make the choice.
This is not a feature we have not reached yet. It is one we do not intend to build. However strong the analysis becomes, analysis is not authority.