Charlot Lab · EFA · Open Weights

The weights are real. Steer them.

This page fetches the Institute's released EFA checkpoints — the exact safetensors files on Hugging Face — parses them in your browser, and runs them on your device in plain JavaScript. No server. No GPU. The flagship is EFA-1: one body-embedding-conditioned trunk that controls three different bodies from a single weights file — swap the embedding, control a different body — with its own energy deciding when one forward pass is enough and when to think harder. The two earlier checkpoints remain as the proof-point lineage.

efa-1 — one trunk, three bodies
efa-hybrid-arm2 — act + verify
efa-flow-arm3 — 3 joints, 1 pass
fetching released weights from Hugging Face…

What you are watching. A chain of coupled pendulums (gravity, damping, neighbor coupling — the exact dynamics in the model's config.json). Gold ghosts mark the goal angles. On the efa-1 tab, ONE downloaded trunk controls whichever body you pick — 1, 2, or 3 joints — by swapping a learned body embedding; the HUD shows the agency gate live: the model's own energy E is checked against the shipped threshold τ each decision, and only when the energy objects does it think harder (K=4). The teal meter is the same potential verifying — scoring its chosen action against a random one at random probe states. The artifact carries its own certificate (exponential stability at every measured attractor + a contraction basin, limits disclosed in the card) and its escalation policy — both live in the config.json this page fetches. Checkpoints are gated and round-trip verified. Honestly small bodies — the point is the architecture identity, and that the artifact you can download is the artifact running here.