WebGPU · loading

This build renders the larger world with WebGPU

Your browser or this context does not expose WebGPU, so the 90k-cell GPU build can't start here.

The WebGL player shows the same engine, same physics, on a smaller world and runs everywhere. Open that instead.

conservation σ (kernel ledger)
0.000000
recognition cost J posted
0.0
eight-tick breath
0 / 8
carried coarse (undrawn)
--
cells resolved 0 / --
cell provenance
Click any cube to read its proof: its rung, the tick it was forced, the J(φ) it cost, and the rule that forced it.
tick 0

The same engine, a far larger world, stepped on the GPU

This is the recognition lattice at depth seven: -- cells the engine was forced to resolve, about eight times the WebGL world. Three.js drives it through WebGPU, and a compute shader steps the entire octree in parallel on every tick, resolving which cells the recognition front has reached and lighting each one as it is written. Nothing here is computed by the page's physics, because the page has none. The cells, the rung of each, the J(φ) each forced cell cost, and σ = 0 are all the engine's, baked into the trace. The GPU does the presentation step, at a scale a per-cell CPU loop cannot keep up with. Click any cube to open its provenance.

undrawn = carried coarse hue = φ-spaced rung bright = recognition demand σ locked at 0
What the GPU does, and what it does not. The compute pass reads the tick each cell was forced (the engine's number) and, per tick, writes which cells are now resolved and their write-flash, for all -- cells at once. It never invents physics. The cost J equals the number of forced interface edges times J(φ) exactly, and across all -- ticks the kernel ledger's σ is 0 to machine precision. Press inject σ imbalance to see the conservation needle leave zero and the run flag, exactly as the kernel falsifier requires.

What is literal here, and what is a scene choice

Literal, proved
The cost law J(φ), the double-entry σ = 0, the descend-iff-forced rule with threshold zero, the eight-children octree from D = 3, and coarse-carry losslessness. The kernel's real output, replayed.
The GPU's job
A WebGPU/TSL compute pass steps presentation: per tick it resolves the recognition front and the write-flash across all cells in parallel. It reads the engine's numbers and never alters them.
A scene choice, named
The demand field, the core, the φ-spaced shells, and the filaments that say where distinctions exist. The engine's response to it is the proven part, not the shape itself.
The next step
This replays a baked trace. The same renderer, pointed at a live server stream of the engine running unbounded, is the only change between this and the full system. The protocol is identical.

Governed by

Generated 2026-06-06T00:00:33. Renderer: Three.js r184 WebGPURenderer + TSL compute. Trace: rfp_trace_large.json, emitted by the real recognition kernel (scripts/cosmogenesis/rfp_emitter.py). A pure Recognition Frame Protocol consumer: every cube traces to a number the engine computed.