Aegis

Aegis-VLA is an adversarial verification and counterfactual world-model harness for sandboxed embodied coding agents ("code as policies"). When an LLM controls a robot by writing Python, two critical failure classes slip past standard test harnesses: (1) Grader reward hacking, where the policy teleports objects via simulator state writes, spoofs exit codes, or uses Python introspection (__globals__, getattr) to bypass rules; and (2) Silent physical failure, where the script exits 0 cleanly, but the robot arm collided with a fragile obstacle or dropped the payload mid-trajectory. Aegis-VLA runs every generated policy through a four-node LangGraph loop (Perception -> Synthesis -> Sandbox -> Auditor -> Repair). It refuses to commit a trajectory until two independent gates pass across 3 domain-randomized physics seeds: 1. Deterministic Physics & Security Telemetry: Multi-layered sandbox isolation (static AST gate, runtime import guard, sys.addaudithook, state-teleportation traps, and process-group kill switches) paired with an HMAC-SHA256 nonce-signed grader report. 2. Semantic Counterfactual Verification: SAM 3.1 instance masks deprojected through the depth buffer into 3D, compared against the 3D goal state and pre-rollout failure forecast predicted by Muse Wonder, alongside an expected goal frame rendered by Muse Image. When either check fails, Muse Spark diagnoses the physical or security root cause (e.g., computing exact Coulomb friction bounds or clearance heights), and the Muse Code CLI agent autonomously patches the Python policy until all 3 seeds pass verification.