# nous-jepa

Repository-local JEPA infrastructure for Phase 176.2: multi-block I-JEPA masks,
online/EMA-target latent prediction, image/video/audio/3D token adapters,
action-conditioned future-latent rollout, multimodal video/audio/text alignment,
the canonical ViT-H/14 configuration, and portable TorchScript encoder export.

The implementations are deliberately configurable so their full published shapes
and small CPU test shapes share the same code path. The Phase 176.2 checklist
remains open because it asks for specific large-dataset training, linear-probe,
zero-shot planning, and downstream robot outcomes; local synthetic tests cannot
establish those empirical results.

```bash
python3 -m pytest tests/ -q
```