peterfriese / system-one-foundation-models

# System One for Apple Foundation Models 🧠⚡️
A lightweight, native Swift 6 bridge integrating **System One decision models** into Apple's **Foundation Models** framework (`LanguageModel`, `LanguageModelExecutor`, `@Generable`).
Evaluate strongly typed `@Generable` structs and enums against application state in **15–150ms** with zero hallucinations, calibrated probabilities, and full Apple Intelligence API compatibility across:
- **On-Device Core ML (`LayaOnDevice`)**: Run Laya's 322M (multilingual mmBERT) and 421M (English/typed-decisions ModernBERT) parameter models locally on the Apple Neural Engine and GPU with zero network calls.
- **Self-Hosted HTTP (`LayaFoundationModels`)**: Connect to `laya-serve` (PR #31 merged into `NandhaKishorM/laya`) speaking the Jev-compatible `POST /v1/systemone` protocol with presets for `localhost:8000`, `localhost:8770`, and hosted `api.impossibl.com`.
- **Cloudflare Clef (`ClefFoundationModels`)**: Run open-weight multimodal decision models—**Clef (27B)** and **Clef-Flash (9B)**—running via Cloudflare Workers AI edge, Cloudflare AI Gateway, or local runner (`Tools/ClefLocalRunner`), judging camera frames and image attachments alongside text in a single feed-forward pass.
- **OpenAI Decisions (`OpenAIFoundationModels`)**: Non-autoregressive decision computing via OpenAI's Decisions API (`POST https://api.openai.com/v1/decisions`) powered by **`gpt-6-luna`** with ~150ms latency, zero output token billing ($0.10/1M input, $0.00 output), and enterprise multi-tenant header isolation (`OpenAI-Organization`, `OpenAI-Project`).
- **TypeSafe AI Cloud (`JevFoundationModels`)**: Full backwards-compatible support for hosted TypeSafe Jev endpoints with automated HTTP retries (`RetryPolicy`), cooperative cancellation, and confidence routing (`RoutingPolicy`).
> [!WARNING]
> **Security Advisory: Never Embed API Keys in Mobile Apps**
> Cloud API keys (`TYPESAFE_API_KEY`) must **never** be hardcoded or bundled inside client-side iOS, iPadOS, watchOS, or visionOS application binaries. Anyone can inspect or decompile mobile apps to extract embedded secrets.
>
> **Safe Deployment Patterns:**
> - **On-Device Core ML (`LayaOnDevice`)**: Run models locally on hardware with 100% offline privacy and zero secrets required.
> - **Backend / Server / CLI**: Use `LayaLanguageModel` or `JevLanguageModel` directly in server-side Swift services or CLI tools where environment variables remain server-side.
> - **Mobile Applications with Cloud APIs**: Route mobile requests through your own authenticated reverse proxy protected by Apple App Attest and Firebase App Check using the built-in `ProxyTransport` (see Mobile Security Guide and Tech Note 0010).
---
## 💡 Why Decision Models in Apple Foundation Models?
Traditional Large Language Models (LLMs) are generative text engines: coercing them into producing deterministic structured decisions requires constrained token sampling or prompt-and-parse pipelines.
**System One decision models** evaluate ty