I started this project with Jev. The live demos now use Clef. There are several options to try without hosting a model yourself:

  • TypeSafe: Jev 1.14 is its current release, available through the System One API. Model details.
  • Cloudflare: Clef-flash is available through Workers AI. Model details.
  • OpenAI: the Decisions API is generally available, currently using GPT-6 Luna. API details.

You give the model data and a question with a set of possible answers. It returns an answer and probabilities. Your app decides what to do with the result.

I’d use this for repeated decisions: which update to show, where to route a request, or whether to flag a change. An agent could gather the information, then use a decision model for that step.

I’m still asking the same question: what extra value does it add? A simple rule may be enough. Price calculations belong in code.

I want to see whether it reduces noise, what it misses, and what it costs to keep running. I still need to check its choices against real examples.

The three demos here are a start. News radar screens news and social topics. Metals Pulse assesses gold and silver price movements. It doesn’t predict prices. Weather Impact matches US weather alerts to everyday activities.

They show Clef-flash’s decisions and estimated model costs. They work, but I haven’t yet measured whether their decisions are better than simple rules. I’ll add more use cases as I find useful ones.