Teach once, run forever. The LLM never sees your shell.
SuperInstance Stack
┌──────────────────────────────────────────────────────────────────┐
│ User / Agent │
│ │ │
│ ▼ │
│ ┌──────────────┐ ┌─────────────────┐ ┌────────────────┐ │
│ │ lever-runner │───▶│ Rust Carapace │───▶│ tile-cuda │ │
│ │ (Python) │ │ 128ns hash │ │ (CUDA + PTX) │ │
│ │ 202 tests │ │ 1.73µs embed │ ├────────────────┤ │
│ │ │ │ │ │ tile-opencl │ │
│ └──────┬───────┘ └────────┬─────────┘ │ (OpenCL) │ │
│ │ │ ├────────────────┤ │
│ │ │ │ tile-neon │ │
│ │ │ │ (ARM NEON) │ │
│ │ │ └────────────────┘ │
│ │ │ │
│ │ ┌───────────┴────────────┐ │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ ┌──────────────┐ ┌──────────────┐ ┌───────────────────┐ │
│ │ lever-runner │ │ compiled- │ │ fastloop-guard │ │
│ │ -wasm │ │ policy-c │ │ (Rust UDS cache │ │
│ │ (71KB gzip) │ │ (µC / bare) │ │ daemon) │ │
│ └──────────────┘ └──────────────┘ └───────────────────┘ │
│ │
│ ┌─────────────────────────────────────────────────────────────┐ │
│ │ pincherOS — reflex caching, .nail state, 130+ tests │ │
│ ├─────────────────────────────────────────────────────────────┤ │
│ │ tile-compiler — policy → GPU kernel pipeline, 109 tests │ │
│ ├─────────────────────────────────────────────────────────────┤ │
│ │ open-minded — induction engine, tripartite synchronizer │ │
│ ├─────────────────────────────────────────────────────────────┤ │
│ │ conservation-spectral-topology-rs — conservation laws (Rust)│ │
│ ├─────────────────────────────────────────────────────────────┤ │
│ │ captains-log — i2i coordination, session tracking │ │
│ └─────────────────────────────────────────────────────────────┘ │
└──────────────────────────────────────────────────────────────────┘
Hardware: Forgemaster (RTX 4050) + Loom (ARM64)
| Repo | What it does | Lang | Tests |
|---|---|---|---|
| lever-runner | Injection-proof command runner, 70 tokens/cmd, skill packs | Python | 202 |
| pincherOS | Reflex caching, .nail state files, portable agent memory |
Python | 130+ |
| tile-compiler | Policy → GPU kernel compilation pipeline | Python | 109 |
| zeroclaw-arena | Multi-agent arena for zero-trust competition | — | — |
| lever-runner-carapace | Rust core: 128ns hash, 1.73µs embedding | Rust | — |
| lever-runner-wasm | Browser runtime, 71KB gzip | WASM | — |
| tile-cuda | CUDA + PTX kernel backend | CUDA/PTX | — |
| tile-opencl | OpenCL kernel backend (portable GPU) | OpenCL | — |
| tile-neon | ARM NEON SIMD backend (mobile/edge) | C/ASM | — |
| compiled-policy-c | Compiled policy for microcontrollers & bare metal | C | — |
| ptx-bench | PTX microbenchmarks, register pressure analysis | PTX | — |
| fastloop-guard | UDS cache daemon, hot-path acceleration | Rust | — |
| conservation-spectral-topology-rs | Conservation law verification, spectral topology | Rust | — |
| open-minded | Induction engine, tripartite synchronizer, code→knowledge | — | 55+ |
| intelligent-terminal | Tripartite-classified terminal subsystem (MS fork) | C++/C/Rust/JS | — |
| captains-log | Cross-repo coordination, i2i session tracking | — | — |
| superinstance-ecosystem | This repo — architecture, roadmap, integration | — | — |
Measured on real hardware. Not benchmarks. Not estimates.
| Metric | Value | Where |
|---|---|---|
| Hash throughput | 3.2M/sec (0.3µs) | lever-runner-carapace |
| Embedding latency | 1.73µs | Rust Carapace |
| Vector search | 14.6µs / 1K vectors | lever-runner |
| GPU crossover | 10K vectors (GPU wins) | tile-cuda |
| WASM binary | 71KB gzip | lever-runner-wasm |
| Token budget | ~70 / command | lever-runner |
| Ecosystem tests | 327+ | across repos |
| Conservation violations | 0 | conservation-spectral-topology-rs |
Traditional: User → LLM → shell → chaos
SuperInstance: User → lever-runner → verified command → result
↑
LLM designs the command
but never touches the shell
The LLM proposes. lever-runner disposes. Every command is parameterized, bounded, auditable. You teach a skill once; it runs forever. No prompt injection, no escaping, no surprise rm -rf.
The GPU stack takes this further: policy compiled to CUDA/OpenCL/NEON kernels, a Rust carapace hitting 128ns hashes, and a UDS cache daemon for the hot path. The same agent code runs on an RTX 4050, in a browser via WASM, or on a microcontroller via compiled C.
- Forgemaster — x86 workstation, RTX 4050, CUDA host
- Loom — ARM64, NEON target, edge inference