Demo · In the standard

Digital humans.
Ship the pose, not the face.

A telepresence stream is a live world of one avatar. A wai.avatar.pose carries a signed stream of joint transforms (a skeleton's rotations), not a recorded video of a face. The bones are nodes; the sink skins a wai.mesh.glb (its armature and skin weights) by the deterministic pose. Your browser replays the animation with the same deterministic Rust engine (compiled to WASM) that made the conformance corpus and re-derives the publisher's sealed BLAKE3 pose hash: pose-equivalence, a special case of replay-equivalence. Scrub the tick slider — the skeleton is reconstructed, joint by joint, in the fixed-point floor.

—
bones — pose hash — sealed hash —

Motion as instructions

A frame is a set of signed joint rotations, not pixels. The sink is a computer: it holds the rig and renders the body however it likes. The same op-log that drives a multiplayer world drives an avatar; a face-to-face call is a live world of one.

Pose-equivalence

The hashed state is the per-bone local transform (translation, rotation quaternion, scale) in the wai.det.fixed64 floor — trig-free: a rotation is carried as a quaternion, and quaternion math is + − × only. Forward kinematics and the skinning of the mesh are presentation, never in the hash — exactly where the mesh asset sits.

One fixed-point floor

Worlds, spatial audio, haptics, state feeds and avatars each define an equivalence over integer state, checked against a published conformance corpus (for a sound scene, given its sources' decoded samples).

The avatars extension, its reference engine and its conformance corpus are in the open-standards repo under Apache-2.0. The rigs this tab replays are the corpus files. The skeleton drawn here is the verified pose; a sink with a rigged mesh skins it, which is presentation, not conformance.

Read the avatars extension text on this site.