Byte-exact, deterministic CPU rendering
The same scene produces the same pixels every time, on every platform we have compared: native and WebAssembly agree bit for bit.
Result, method and evidence
Result: for eight scripted workloads (settings screen, hover, typing, scrolling chat, fluid smoke, theme cross-fade, charts, cinematic effects) at two sizes, 976 of 976 frames hash identically between the native renderer and two WebAssembly builds (scalar and simd128). The first 1920x1080 charts frame has the same SHA-256 on all three (8023bc0c...f6b4). Separately, 125 PNGs from 18 of the repository's example programs were rendered twice and compared: 0 differences.
Method: every frame is hashed (FNV-1a-64 of the RGBA buffer) and compared against the native oracle, run under Node/V8; the spot check compares raw RGBA SHA-256.
Scope: CPU rendering with the shipped fonts; this is not a claim about GPUs or about other CPU architectures than the ones tested. Measured 2026-11-11.
Check it Licensed source customers: run web/tools/check-wasm-exact.sh. Binary customers: render the same scene with the native runtime and the wasm module and compare SHA-256 of the RGBA output (see WebAssembly guide).
Evidence in the repository
web/tools/check-wasm-exact.mjsframe-by-frame comparison against the native oracleweb/README.mdthe byte-exactness check and its limitssite/content/gallery/proof/byte-exactness.jsonreview branchthe proof record produced for this sitedocs/1.0/PERFORMANCE-RELEASE-20261105.md32 PNGs byte-identical to the qualified baseline after the performance release
