Tandem: the bimanual VLA that knows to refuse

Tandem is a bimanual manipulation stack: two simulated SO-101 arms in MuJoCo set a dinner table from a spoken instruction. Speechmatics transcribes the voice live. A Qwen2.5-1.5B model, quantized to INT4 and run locally through OpenVINO GenAI, only parses the intent; deterministic code turns that into a plan from the live scene state. Before either arm moves, a ~200-line deterministic gate checks every step for reachability, gripper state, object presence and collision, then allows it, repairs it, or refuses it with a sentence saying why. No model takes part in that decision. Neither arm can reach both the props and the place setting, so objects pass from one arm to the other through a shared zone. Grasps close on measured contact, not a preset width. Every episode is seeded and domain-randomized. Every number is regenerated by scripts in the repo. The gate scores 24/24 on a set that includes 8 must-allow cases, so refusing everything would fail. Over 20 randomized seeds: drawer opened 100%, plate placed 90%, fork 90%, spoon 65%, mug 55%. Live speech: 922 ms median final-transcript latency, with barge-in that aborts a running plan. A learned perception net replaces simulator state at 23.8 mm median error, and the gate reaches the same verdict from it as from ground truth on 94.1% of 456 steps. The nets export to OpenVINO at FP32, FP16 and NNCF INT8, benchmarked. What does not work is published too. Water lands in the cup in 0 of 20 seeds: the carton empties in 55%, but the stream misses. The distilled visuomotor policy makes 0 of 18 picks against the scripted expert's 14. Benchmarks ran on an AMD CPU, not Intel hardware, and every report says so. MOCKS.md and CLAIMS.md draw the line between measured and simulated. Reviewer path: JUDGES.md grades the gate in about a second, with no API key, GPU or model download.