Every company should own the intelligence it pays for.

Most companies rent their intelligence. Every day their agents send thousands of calls to a frontier model. Each one gets an answer, and each one is paid for. Then the transcript, the tool calls and the human who fixed the answer all expire in a log bucket. Nothing is kept.

That is a strange way to treat the most valuable data a company produces. It is a record of real work, on real systems, with real outcomes attached — exactly what you would want to teach a model with.

The L1 engineer idea

Nobody hires an L1 engineer for what they know on day one. They get good at specific tasks by doing them next to someone better, with a lead checking the work. After enough reps, they own the queue.

We train open models the same way. The frontier model is the senior. Verifiers — checks written against your own systems — are the lead. A small model learns the twenty tasks your agents repeat every day until it matches the senior on those tasks, and the senior stays on call for everything else.

Why the golden dataset matters

The model is only half of what you end up with. The other half is the golden dataset: every example checked before it is admitted, versioned, and yours. Models will keep changing. A verified record of how your company does its work outlives all of them — when a better open model ships, you retrain instead of starting over.