Four releases, and they don't agree on the same design — or even the same year. Physical Intelligence's π0.7 (April 2026, arXiv:2604.15483) steers one 5B policy — a 4B vision-language backbone, a video-history encoder, and an 860M-parameter action expert — by conditioning on four channels at once: the language command, a running language-coaching stream, episode metadata about how a demonstration was performed, and a visual subgoal image. Physical Intelligence's own framing is careful: this buys 'early,' 'first signs' of compositional generalization — skills recombining onto tasks with no matching demonstrations — not a solved problem. NVIDIA's GR00T N1.7 fuses differently: it replaces its vision-language backbone with Cosmos-Reason2, a model built for step-by-step reasoning, feeding the same flow-matching action head, and reports the reasoning backbone improving long-horizon task decomposition and cross-embodiment transfer. Alibaba's WorldVLA (2025, arXiv:2506.21539) goes further still — one autoregressive transformer generates both the next action tokens and the next image tokens from a shared vocabulary. That's a real fusion, not a world model bolted beside a policy: the world-model half conditions on the action model's image understanding to predict the future, the action half conditions on the world model's frame prediction to act, and the paper reports each improving the other. Figure's Helix (2025) is the honest counter-example: no world model at all. A slow 7–9Hz vision-language system hands a semantic vector to a fast 200Hz reactive policy, split by timescale, nothing predicts a future frame — and it runs on real humanoids today. The field has not converged. You should know the shape of the disagreement, not just one side of it.