lejudge-jev-jepa
Natural-language constraints for JEPA world-model planning, judged by a decision model instead of an LLM.
How it uses Jev
1. Imagine. LeWM's predictor rolls out 300 candidate action sequences in latent space (CEM, 30 iterations). 2. Describe. Linear probes turn every imagined latent into a handful of words from a closed vocabulary (grid cell, wall contact, angle bin, contact, speed). The judge never sees a coordinate. 3. Judge. Jev, TypeSafe's System One decision model, answers one typed yes/no question per (constraint, step), "At step…
- Languages
- Python 82% · TeX 18% · Makefile 1%
- License
- MIT
- Last activity
- Sep 2026
- Added
- 2026-09-23
- Stars at snapshot
- 9
- Status
- Community listing
Source: community catalog. Details and metrics reflect the published community snapshot and are not a performance endorsement. Documentation details extracted from GitHub · the README.