Runs in your browser
Each demo is a standalone page with its own design, running in your browser. The GPU demos use WebGPU; the fractal explorer falls back to WebGL2.
WebGPU + CPU
One integer kernel (add, subtract and shift only) runs on your GPU and your CPU in the same tab and is stitched into one image. Identical bytes give an invisible seam, an all-black XOR diff and a matching SHA-256 seal; flip one bit and all three break.
WebGPU / WebGL2
Mandelbrot, Julia, Burning Ship and Tricorn with smooth zoom and pan, rendered on your GPU. The repository holds the MIND sources for the complex arithmetic, iteration and colour palettes.
WebGPU
A particle simulation of MIND's cognitive cycle: sense, think, act, verify, learn. A SHA-256 hash chain makes each seeded run tamper-evident and replayable.
Simulation
In progressFollow one token through a model-serving pipeline written in MIND: paged KV cache, FlashAttention-2-style attention, speculative decoding, continuous batching and disaggregated prefill/decode, behind a governance gate.
Governance simulation
Launch ten simulated attacks against MIND-Law's nine invariants and six extensions (forge a constraint, cut a feed, replay evidence, collude) and watch every refusal sealed as an evidence record. MIND-Law is STARGA's implementation of the 512 constraint by Dr. J.M. Watson. 512 defines seven canonical conditions. I1–I9 are MIND-Law's implementation of them. Extensions are MIND-Law's own and are not part of 512.
Code governance
Commercial · v0.2.0Nine Q16.16 fixed-point metrics over a typed dependency graph, with architectural rules declared in .mind and enforced fail-closed. Session deltas are chained with HMAC-SHA256. The demo reads a dated, checked-in scan of MIND-Mem, not a live feed.
WebGPU benchmark
A register-tiled WGSL GEMM shader against ONNX Runtime Web's WebGPU backend on the same matrices, on your own GPU. The number you see is your measurement, not ours.
From the cards
Each figure is the project's own published result, repeated from its card.
10.5× to 50.5×
GPU Path Tracer: lower wall-clock than four reference renderersMeasured48.4×
GPU Path Tracer: steady-state per frame against Mitsuba 3, same GPU96.06%
MIND-Nerve: top-5 accuracy across 11,922 routing candidatesMeasured107
MIND-Mem: MCP tools, one workspace serves 19 coding clients15 · 10
MIND-Law War Room: invariants and simulated attacksOpen repositories
Open repositories you can clone, build and benchmark yourself.
GPU computing
A path tracer written in MIND with an optional BVH-accelerated CUDA backend. Its harness times it against Mitsuba 3, Falcor, Cycles and LuxCore on the same scenes, with the method and raw results in the repository.
Game engine
A chess engine written in MIND: hybrid alpha-beta and Monte Carlo tree search with NNUE evaluation, Syzygy endgame tablebases and the UCI protocol. Batched GPU evaluation runs through the MIND runtime.
SI memory
PyPI 5.0.3Replayable memory for SI coding agents. Every write goes through propose, review and apply; every applied decision is hash-anchored in a Q16.16 audit chain that replays byte-identically with query, corpus, configuration and scoring date held fixed.
Intent router
PyPI 0.3.1 · Apache-2.0A small classifier between the request and the agent runtime returns only the top-K relevant skills, tools and MCP servers, so library size decouples from token cost. The encoder is compiled to native Q16.16 fixed-point.
Send us your project. We feature work from the community here and in the newsletter.
