The short version
- AI models build most of Bench, the same way agents do work for you.
- The model that writes a feature never gets to approve it. A separate model checks it.
- Money, legal, and checkout never skip that check.
- A failed check can't be overwritten by a person saying it looks fine.
How a feature gets built here
Say we're adding a new step to the checkout page. One model writes the plan. A second model builds it. Then a third model that had nothing to do with either of those reads the result and tries to find what's wrong with it. It doesn't get told the work is good. It gets told to break it. If it finds a problem, the fix goes back through the same check. Nobody on our side gets to type "looks fine" over a failed check, and that includes us.
Why we do it this way
A model grading its own homework will grade it well. It wrote the thing, so it reads the thing the way it meant it, and it misses what it never thought of. Every model we've used does it. A second model with no stake in the answer is the cheapest honest reviewer we've found.
Where it's required, no exceptions
Anything that touches your money, your legal agreements, or the page where you pay us. Anything that could be forged if a rule were wrong. Those never ship on one model's word. Smaller changes, like fixing a sentence on this blog, don't need three models, and we don't pretend they do.
What it means for you
When a Bench update lands in your office, at least two independent models have already argued about it, and the one that wrote it lost every argument it was going to lose before you saw the screen. That's what "done" means here.
Questions
Why more than one model?
Because a model reviewing its own work approves it. A second model with no stake in the answer finds what the first one missed.
Does everything go through this?
Anything touching money, agreements, or checkout always does. A one-line copy fix doesn't need three models, so it doesn't get them.
Can someone override a failed check?
No. If the independent check fails, the work goes back. Nobody writes over the fail.
Keep going
- A chatbot thinks. A harness gets a job done.
- You get a web page first. The PDF comes after you say it's done.
- When we fix a bug, we also block every bug like it.
Every Bench word above is on Glossary / FAQs, in plain English.