Jev, Laya and the case for smaller AI decisions

Picture a support inbox. A message arrives, and someone needs to decide whether it is urgent, routine or unclear. A chatbot could write a polished paragraph about it. But the useful output may be much smaller: a category, a reason to check it, or a clear “I do not know.”

That is the question behind our early SLOP research: can small, well-defined decisions be more useful than more generated text? SLOP is an internal research project, not a released product or a claim that we have solved the problem.

Why Jev and Laya caught our attention

TypeSafe describes Jev as a model that answers typed questions about a given state. Its published interface includes choosing an option and assigning a score. The open-source Laya project explores related typed decisions. Both are useful reference points for how software might ask a narrow question and receive an answer it can act on. They are not evidence that either model is best for our tasks.

The practical test is less glamorous than a demo: ask the same question on a fixed set of examples, record the right answers in advance, compare against simple rules, and count the mistakes. When the answer is uncertain, the system should be able to defer to a person rather than sounding confident.

Where Bonsai fits

We are also testing a local Bonsai-compatible route for extracting information from research papers. That is groundwork for a research workflow, not a benchmark victory or a deployed customer system. Keeping source material and evidence close to the experiment should make it easier to check what a model actually used.

What we still need to learn

We have not published a controlled Jev-versus-Laya comparison on our own tasks. Before making a performance claim, we need a frozen dataset, agreed scoring rules and reproducible runs. A plain rules engine may win on some questions. That would be useful to know.

The aim is not to replace judgment with a magic score. It is to learn where a small answer helps someone move forward, where it fails, and when the honest next step is to ask a human.