aien-spark/aien-sovereign-core

The earlier **Linux-hosted** AIEN runtime, written in Rust: agent CLI, persistent memory (Cortex), a unified-memory KV-cache with copy-on-write branching, a continuous-batching scheduler, and an inference ABI whose GPU path runs AIEN's own Omega kernels. It runs on top of Linux on the NVIDIA DGX Spark (Grace Blackwell GB10) and is the reference for what the native stack must beat.
**This repository is legacy.** AIEN's own stack (the aienos kernel, the omega runtime, physics FORGE) is replacing it. Under ADR 0024, working Rust here is kept as temporary scaffolding, not rewritten into C: it stays behind language-neutral boundaries and Omega replaces components one at a time after passing the same conformance tests. New Rust is allowed where it is not hardware-bound (for example the DEV-MODEL-0 program, a local development model on AIEN-owned inference). Component ownership (for example the Cortex now owned by omega, per ADR 0022) is recorded in aien-architecture; here crates/cortex-rs is non-authoritative and to be retired, not ported.
No Python in this repo (a CI guard, scripts/check-no-python.sh, fails the build if any Python file is added), and no comparison with other people's systems: speed is measured against the machine's physical limit, and correctness against our own reference.
> SUPERSEDED 2026-10-01 by ADR 0024. Previous text, kept as history: **This repository is legacy.** The project's target language is C, and AIEN's own stack (the aienos kernel, the omega runtime, physics FORGE) is replacing it. No new Rust is added here, with one declared exception: the DEV-MODEL-0 program (a local development model on AIEN-owned inference) may temporarily extend this Rust substrate while the C replacements are built. Component ownership (for example the Cortex now owned by omega, per ADR 0022) is recorded in aien-architecture; here crates/cortex-rs is non-authoritative and to be retired, not ported.
Current state
Research-grade and pre-alpha. Crates build and have unit tests on the DGX Spark; execution is CPU plus GPU through AIEN's own Omega kernels. Modular MAX and Mojo were removed from the project on 2026-10-15. Nothing here is qualified on hardware as a finished system. Live status and sequencing: CURRENT_EXECUTION_PLAN.md and the implementation status snapshot. Component detail and platform status: docs/PLATFORM_MATRIX.md, docs/AIEN_RUNTIME_ARCHITECTURE.md.
Known limits: the GPU fallback counter covers only instrumented operations, so execution is not exclusively accelerated; Qwen full-runtime qualification is not complete; the C1 evidence campaign is in progress.
Measured results
Every headline performance number must resolve to a reproducible command and an artifact bundle (commit identity, hardware and environment record, exact command, raw samples, SHA-256 digests, measurement definition, reproducibility steps). Earlier figures (branch fork latency, COW page mutation latency, memory sharing ratios, service comparison table) are withdrawn until regenerated to that standard. The evidence bundles live in aien-spark/benchmarks.
Standing rules
Language rule: Rust is scaffolding, Omega is the destination, and C or assembly stay only where hardware, boot, ABI or measurement justifies them (ADR 0024, which supersedes the old "C is the target, no new Rust" rule). No Python anywhere in repos, build, CI or tooling. No CUDA toolkit and no new CUDA dependence. No systemd in AIENOS. No outside dependencies in the trusted base. No Modular MAX or Mojo anywhere in the tree (removed 2026-10-15); kernels are Omega or Rust, closest to the metal, fastest wins, build what is missing.