A growing share of AI models are small classifiers that return probabilities instead of generated chats and text: a yes/no probability, a selection from a bucket list, or a scoring scale.
Barndoor AI Gateway now supports decision models natively, bringing this model category under the same governance as everything else, including TypeSafe’s System 1 model, Jev.
Decision models, also called System 1, are high on the hypetrain, but what are they most useful for? LLMs generate text, code, and any other long-form information format. System 1 models (i.e. classifiers) will just give you a bucket. Is it yes or no? Is it a clothing or technology query? Is it a good answer or a bad answer?
It’s not quite the swiss-army knife that LLMs provide. Yet specific applications are well suited to it, such as model routing.
How it works
Jev plugs into the gateway two ways.
First, you can call it directly as a model (the same way you’d call any chat model) with the same policy enforcement and logging. The route is POST /v1/systemone, which mirrors TypeSafe’s path and request/response format, so if a team is already using TypeSafe SDKs, they can just change the base URL. The models available are jev-latest and jev-1.13.0.
Second, Jev can serve as the decision-maker inside a model routing policy. So instead of asking a chat model to figure out which rule a request matches, the gateway asks Jev, which is built for a yes/no classification response.
System 1 models applied to routing
Model routing is based on the assumption that there is a most optimal model for a given prompt. The agent asks through the LLM endpoint, “write an email!” and it should do it with a model with less intensive usage. The agent uses the same endpoint and asks, “refactor the backend!” and it should access the biggest model it has.
Before shipping compatibility with System 1, the Barndoor policy router used an LLM-as-judge approach to answering this problem. An LLM selected by the administrator sits on top of the request and uses policy information to figure out how and when to funnel requests to different models.
We then swapped in Jev for the same task and scored both approaches two ways — decision quality, judged independently by two separate models, and raw performance based on latency and cost per decision.
The result: Jev matched a Claude Haiku model’s pick quality, while running roughly 7x faster at a fraction of the cost.
Why it matters
Because this runs through Barndoor’s policy routing connection, any classifier model built to the same API can be routed the same way. Every call, direct or routing, flows through the same data protection checks and logs the same usage and cost as the rest of your AI traffic.
Most governance platforms aren’t ready for this type of decision model. The Barndoor AI Gateway provides the same policy, spend tracking, and audit trail that already covers your mainstream models and LLM usage. To learn more, take a product tour or schedule a demo.
