What 73% Actually Means
Everyone building AI products right now is routing every question through a language model. Every. Single. One.
That's like calling a doctor every time you have a headache. Most headaches, you know the answer — drink water, take ibuprofen, get some sleep. You don't need a doctor. You need to check what you already know.
That's what our system does. When a question comes in, it checks: do we already know this? Is there a cached answer? Can we match the structure of the question to a known capability without any guessing?
Seventy-three percent of the time, the answer is yes. The question gets routed to the right place by structure — not by probability — and the answer comes back in milliseconds for free.
The other twenty-seven percent? Those are the genuinely novel questions. The ones where you actually need a model to think. And for those, we use one. We're not allergic to language models. We just don't think you should need one for "what time is it in Tokyo."
The number isn't the point. The principle is: calculate when you can, guess only when you must. It's not a breakthrough. It's just discipline.