lejudge-jev-jepa

What it is

A PushT study and cost module for LeWorldModel planning. It turns imagined latents into closed-vocabulary facts, asks Jev typed questions about natural-language constraints, and adds the probabilities to a CEM planner's cost. Ships oracle and keyword baselines, cached responses, and an offline paper rebuild.

How it uses Jev

Jev answers one typed yes/no question per (constraint, step) about probe facts from imagined latents, returning P(true). Python folds these into the CEM cost: max over steps for never, 1−min for always, expected rubric level for soft. Judged step-facts are memoised across candidates, iterations, replans, episodes and runs.

Primitives:choicescorenoul

Technique worth stealing

Judge closed-vocabulary probe facts with a decision model instead of an LLM, and memoise judged step-facts to collapse a 300-candidate population to a few hundred calls.

Try it

See README's 'Reproduce everything' and 'Use it in your own planner' sections.

View on GitHub