diff --git a/README.md b/README.md
index 6ca4109..c8492cb 100644
--- a/README.md
+++ b/README.md
@@ -456,7 +456,7 @@ Every response surfaces telemetry from each enabled wedge (`guard_clamped`, `gua
 
 ## What Tether is and isn't
 
-**Is:** the deployment layer between a trained VLA and a real robot. Cross-framework export verified at cos=+1.0000000 on six VLA families — SmolVLA + pi0 + pi0.5 (flow-matching, num_steps=10) + GR00T N1.6 (DDPM DiT, num_steps=4, **with Eagle 2.5 VL backbone producing live image+language KV**) + DreamZero (world-action model, joint video + action diffusion) + OpenVLA (shim) — plus a composable runtime (serve + safety + turbo + split), edge-first design targeting Jetson + desktop NVIDIA GPUs.
+**Is:** the deployment layer between a trained VLA and a real robot. Cross-framework export verified at cos=+1.0000000 on four VLA families — SmolVLA + pi0 + pi0.5 (flow-matching, num_steps=10) + GR00T N1.6 (DDPM DiT, num_steps=4, **with Eagle 2.5 VL backbone producing live image+language KV**). DreamZero (world-action model — config + PyTorch runtime today, DiT ONNX in progress) and OpenVLA (optimum-cli shim + postprocess helper) are supported but not yet numerically verified through a Tether ONNX export. Plus a composable runtime (serve + safety + turbo + split), edge-first design targeting Jetson + desktop NVIDIA GPUs.
 
 **Isn't:** a training framework (PyTorch/JAX own that) or a cloud inference provider (vLLM/Baseten own that). Tether's moat is the deployment toolchain: cross-framework ONNX with verified numerical parity, composable safety wedges, ROS2 + Docker + HTTP serving, and a deterministic export receipt (`VERIFICATION.md`) your QA team can audit.