Code-JEPA is a research project on hard-negative-sensitive JEPA representations for code, later used for reranking, RLJF-style preferences, hindsight agent training, and failure-memory exploration.
Start here:
Project.md— scientific source of truth.Work.md— active work and one next step.CLAUDE.md— agent rules for this repo.docs/diagrams/code-jepa-project-map.pdf— 1-page technical map.docs/design-notes/data-prep-pipeline.md— current tokenizer-agnostic prep pipeline.