Jev Integration with Werkfaden

Sep 19

Jev is a new component integrated into Werkfaden that provides typed probabilistic decisions about project state, rather than generating text.

Had some time to wire something fairly new into Werkfaden this week: Jev. And this one is interesting because it is not another LLM sitting next to the coding model and commenting on its work. Jev does not generate text at all. The surrounding workflow defines the questions and the possible outputs. Jev gets project state and returns typed probabilistic decisions about it. Which happens to fit WF rather neatly. The coding model still does the messy part: understand the bug, form a hypothesis, query the project state and make the patch. Werkfaden still handles retrieval and authority. The new bit is that the resulting findings can now be decomposed into small explicit decisions and run through Jev before the next iteration. So instead of asking another LLM: “does this patch look right?” the loop can ask things closer to: does this exact architectural condition hold? does the current evidence support it? did this particular property improve after the patch? and get probabilities back rather than another paragraph. The screenshot is from one of the first runs. One change relaxed an old single-channel restriction. The expected N-channel binding signal moved 0.31 → 0.43. At the same time the broader multichannel architecture score moved 0.35 → 0.32 and the scientific-null dimension 0.49 → 0.39. So: the patch moved the thing it was supposed to move, while the surrounding state still said nope, not done yet. Good. The interesting part is the division of labour: LLM → reason + change Werkfaden → retrieve project state code → define the decision graph Jev → return calibrated decisions Then run it again. Still very early. But this is getting closer to what I wanted from WF in the first place: not another model producing more prose, but a tighter loop where assumptions can be turned into explicit decisions and checked against the current project state before they quietly become implementation. Keep the thread. Verify the change. Well done @typesafeai . #AIEngineering #LLMOps #Jev

Original post