We Built Compiled AI. Jev-Like System 1 Models Belong Inside It.

The agent world spent two years building systems that think out loud. Every step was another prompt, and every failure was a paragraph nobody could audit.

We bet the other way. Compile the workflow once, then run it as code. The model stays on a leash, with bounded tools, fixed schemas, and no open-ended control loop. Intelligence goes in at design time, and reliability comes out at run time. In healthcare operations, that's how you get through compliance.

Now a new class of model is arriving that fits this approach: Jev-like System 1 decision models. They don't write essays. They answer typed questions (pick one, give a score, yes or no) with a calibrated confidence. The industry is starting to see what we built for from day one: decisions are not prose.

System 1 models don't replace Compiled AI. They slot into it.

In our architecture, decision models become compile-time components. The compiler lays out the decision points: what can be approved, what must escalate, which tools are allowed, and how confident the model has to be before anything happens. A System 1 model fills that point. The workflow stays deterministic and versioned. Generative models still handle the messy extraction, decision models check, and code acts.

They can also watch from above. Put a decision model alongside a compiled flow as an observer, scoring each step, flagging drift, and marking which runs are clean enough to auto-approve and which need a human. The run never re-plans in English. It stays code, now with a typed record of every judgement, tied to a model version and a line of policy rather than a system prompt.

That's the edge. Chat agents rent every new model and hope the next prompt holds. We build the model into the factory, behind stable contracts, with golden test cases and swappable evaluators. When a better decision model ships, we recompile the gate instead of rebuilding the agent.

One thing doesn't change. We explore everything that could move the needle, and System 1 models are high on that list. But nothing reaches a customer workflow until it has proven itself on our own test cases and cleared our compliance bar. We explore fast, and we ship carefully.

Don't prompt the future. Compile it.

— XY.AI Labs