adapterOS

Deterministic ML inference platform for Apple Silicon

adapterOS is a Rust-based ML inference platform optimized for Apple Silicon. It provides K-sparse LoRA routing, Metal-optimized kernels, and policy enforcement for production and air-gapped deployments.

---

What is adapterOS?

adapterOS enables deterministic multi-adapter inference on Apple Silicon:

• K-Sparse LoRA Routing — Dynamic adapter selection with Q15 quantized gates
• Deterministic Execution — Reproducible outputs via HKDF seeding and canonical serialization
• Policy Enforcement — Runtime validation with canonical policy packs
• Zero Network Egress — Air-gapped serving over Unix domain sockets
• Multi-Backend — MLX (primary), CoreML/ANE, Metal kernels

---

Quick Start

# Bootstrap CLI launcher (builds/rebuilds aosctl as needed)
./aosctl --rebuild --help

# Initialize
./aosctl db migrate

# Start
./start

See QUICKSTART.md for full setup.

---

QA Visual Gate

The canonical UI quality gate is Playwright-based and runs as dual-browser blocking checks on macOS:

cd tests/playwright
npm run test:gate:quality -- --project=chromium
npm run test:gate:quality -- --project=webkit

Baseline policy:

• Canonical visual baselines are macOS snapshots (*-darwin.png).
• Gate execution enforces a snapshot contract precheck that fails on missing active or orphaned baselines.

See tests/playwright/README.md for lane composition, run artifacts, and troubleshooting.

---

Requirements

• macOS 13.0+ with Apple Silicon (M1/M2/M3/M4)
• Rust (stable, see rust-toolchain.toml)
• MLX — brew install mlx (for inference/training)

---

Documentation

Document	Description
QUICKSTART.md	Get running in minutes
ARCHITECTURE.md	System design and components
DETERMINISM.md	Reproducibility and replay
POLICIES.md	Policy engine and packs
docs/	Full documentation index
docs/governance/	Governance policy and operations

---

License

Proprietary. Copyright © MLNavigator Inc R&D | James KC Auchterlonie. All rights reserved.