Dunes
GPU displacement paired with CPU contact height, analytic slope, and skiff attitude.
Integrated field application
Recover a stranded survey core and carry it across a flooded dune basin to the authored landing pad. The terrain, swell, wind, and moving breach body are not backdrop: CPU game rules query the canonical fields while the GPU consumes a named, assayed materialization.
Awaiting expedition start
C recenters the chase camera · R retries · P pauses · F fullscreen. Changing lanes reloads the same mission; fault injection is a teaching mode, not a competitor benchmark.
The recorder distinguishes authored-field arithmetic from mesh interpolation and simulation behavior. “Zero readbacks” means gameplay never waits for a GPU result; it does not claim that rasterized triangles are identical to the continuous CPU field.
Application map
This is not a benchmark dressed as a game. It is a compact application pressure test: sparse authoritative queries on the CPU, dense visual materialization on the GPU, no per-frame readback, and an explicit fallback path.
GPU displacement paired with CPU contact height, analytic slope, and skiff attitude.
GPU water displacement; CPU four-probe buoyancy at the same game clock.
GPU flow tracers; CPU force continuously changes the route and handling.
GPU body animation; CPU proximity and collision against the authored spine and radius.
The default WebGPU lane and the WebGL2 fallback materialize the authoritative adapter-authored tree through native TSL. The optional TypeGPU lane consumes the precompiled matter-field/typegpu/three integration: a separately maintained port checked against the canonical fields, not generated from their source. CPU gameplay remains the number/dual authority in every lane. Device receipts measure finite samples and declared absolute budgets; they are evidence, not a universal floating-point proof.