Team Interp Boys

EmbodiedCI is an agent harness that debugs robot world models from the inside, fixes them, and only ships a fix that passes a strict gate. The model under test is Meta's V-JEPA 2-AC (1B-parameter ViT-g encoder, frozen) driving a Franka Panda arm in the robosuite simulator on a cube-lifting task. We test it on clean seeds and five drifts it never trained on: low friction, heavy cube, dim light, a blue cube and a moved camera. Raw, it succeeds safely in 30 of 40 episodes, and with the camera moved it fails every time. The harness runs Muse Spark agents in a recursive-language-model loop. Each agent's only tool is a persistent Python REPL. Episodes, encoder activations and probe results live in REPL variables outside the prompt, so the agent writes code to read just what it needs. It can spawn sub-agents and works on a fixed budget of turns, tokens and time. Interpretability tools are plain functions in that REPL. Linear probes on encoder block 19 read where the model thinks the cube and gripper are. Attribution (x·∂L/∂x) ranks the features behind a bad action. A failure-ahead detector built on probe confidence and JEPA surprise warns about a step early (AUROC about 0.93). rollout(), try_guardrail() and try_head() test any hypothesis or fix in simulation on an H100. The agents checked 6 hypotheses against the logs (5 confirmed, 1 refuted) and found two different bugs. On the blue cube, cube-probe confidence drops from 0.83 to under 0.5, which is a perception bug. With the camera moved, the cube reading stays sharp but the gripper estimate is off by 24 to 30 px, so the arm climbs too fast, which is a control bug. Fixes come from inside the model (probe recalibration, DAgger on its own mistakes, circuit-restricted fine-tunes) and outside it (guardrail code). Every candidate faces a gate on unseen seeds: nominal success can't drop, unsafe actions can't rise, it can block at most 5% of actions, and it can't break an episode that used to pass. Four of five candidates were rejected, including two guardrails the agents wrote themselves, which blocked over a third of all moves. Result on held-out seeds: safe success went from 78% to 100%. On the same 30 unseen seeds, v0 got 16/30 and the champion got 30/30, with 14 fixed and 0 broken. Blue cube and moved camera both went from 0/6 to 6/6. We trained about 1,900 parameters with about 5,000 simulator steps on one H100 for about $7. The robosuite SAC benchmark trains for around 250,000 steps on this task. Live studio: https://embodied-ci.vercel.app