Two machines and an argument
Tenuit designs its screens the way a newsroom checks a story: one agent builds, a second agent tries to knock it down, and a person rules on the disputes.
Most AI-made interfaces fail the same way. They look finished and fall apart on contact with real data: a count that disagrees with the table beside it, a date that lands on a Sunday, a button that promises something the system can’t do. The design loop exists to catch that before an engineer writes a line of code.
Why a second agent
A model grading its own work is generous. A separate grader, briefed only with the criteria and the shared facts, is not. It writes its own tests rather than trusting the builder’s, and says what it could not check. When builder and grader disagree about a fact, a person rules once. The ruling goes into a checker, so the same argument cannot come back.
What it caught
- Two counts, two truths. A header said three items needed you; the list beneath counted different rows. Fixed with one rule both read from.
- A date that drifted. One household’s last review read Jun 26 on one page and Jun 24 on another. The checker now fails it.
- A label that flattered. Two skills were marked “General” while reading the firm’s data. Relabelled by what they actually read.
- A success that wasn’t. A share page told an outside reader “All 15 skills added”. They had no account. The button was removed.
Two brands, one loop
Tenuit’s look is for the advisor at work. Anything a client sees wears the firm’s own brand: its mark, type, colours and voice. The loop grades both, and a client piece fails if it looks like Tenuit.
Where the person stays
A person signs the criteria, picks between directions, and rules on anything that touches compliance or real client data. Everything reversible, the loop decides for itself.