Dómnir: Runes, Robots & Ragnarök

What happens when AI confidently writes code for a flying robot? We don't hand it the keys to the sky. We give it a sandbox, a physics engine, and an adversary named Ragnarök. Dómnir is a mythology-inspired, containment-first platform for testing AI-generated quadrotor controllers. Using Vultr Serverless Inference, it generates Python control code and executes untrusted controllers inside Microsandbox microVMs. A trusted MuJoCo simulation processes bounded control commands, measures physical trajectories, and evaluates mission completion separately from safety constraints. In the Ragnarök Arena, differential evolution, Sobol sampling, and random-search baselines probe for failures within explicit budgets. The 3D Gauntlet contains takeoff, slalom navigation, gate transit, pad approach, and precision landing. Our trusted baseline completed all five stages on time. AI-generated controllers can run through the isolated evaluation path, but we do not claim they have completed the entire Gauntlet. Dómnir also builds signed Court Records with controller and world hashes, contact and stage evidence, and deterministic replay checks. Change the recorded result and the verifier rejects it. Isolation and cleanup are reported explicitly, rather than hidden behind reassuring green badges. Built for C1: Blast Radius Zero, Dómnir explores a simple idea: before AI-written code controls real machines, can we break it safely and verify what actually happened?