kairos-press-button-v16-2-vjepa2

Kairos press-button v16.2 — V-JEPA 2

Little-WAM/Kairos visual-action diffusion experiment on 50 press-button demonstrations; 14-dimensional absolute joint actions, horizon 16.

Steps are phase-local. P2 training is complete. The final checkpoint is p2/step_007000.pt; the earlier step-5,500 checkpoint is also retained. P2 measures surprise with four denoising steps and a running 90th-percentile write gate. P3 is not included.

The frozen external encoder is facebook/vjepa2-vitl-fpc64-256, revision b3c1679b7c34d3255ef3547f27c7b226aefab26f; its weights are not included. Each causal 16-frame clip is resized to 256×256 and encoded jointly. Temporal mean pooling and spatial adaptive averaging produce 10×12 tokens with 1,024 channels; this records the trained geometry.

Each phase contains model weights and JSON configuration/normalization metadata. .pt files retain the native dit state dictionary and model contract; optimizer/trainer state, feature caches, logs, and .conf files are excluded. latest.json resolves weights relative to its phase folder. Original training paths in train_args.json require local remapping.

Use Little-WAM branch research-marina/v16.2-surprise-memory-slots-vjepa2, commit 798465279d1a7110a886d3101bea146c47106f57. These research checkpoints have no RGB decoder or supported streaming inference path; the encoder-specific restriction overrides inherited generic inference metadata.