AGENTIC SYSTEM 1 · DUAL-BRAIN ARCHITECTURE · docs
design-os-system-one: Knowledge Map & Documentation Index
Welcome to the comprehensive knowledge base for design-os-system-one and the System 1 Dual-Brain Architecture. This documentation synthesizes on-device System 1 intelligence (Laya-MLX) with cloud-scale triage (TypeSafe JEV) to build deterministic, sub-10ms autonomous agent pipelines.
🧭 Learning & Exploration Roadmap
| Order | Document | Focus Areas | Target Audience |
|---|---|---|---|
| 01 | 01. System One Philosophy | System 1 vs System 2, Calibrated Probabilities, and the “Code in Control” paradigm. | Architects, New Developers |
| 02 | 02. Core Primitives | The 3 core questions: Choice, Score, Noul, and multi-question atomic batching. |
Backend Developers, AI Engineers |
| 03 | 03. State & Context Management | Structured State (JSON, Text, Tables, Code) and context window optimization. | Prompt Engineers, Backend Devs |
| 04 | 04. Confidence & Gated Routing | Distinguishing probability vs confidence, confidence-gated routing, self-consistency. |
Reliability Engineers, Tech Leads |
| 05 | 05. REST API Reference | HTTP POST /v1/systemone, Auth headers, JSON schemas, status codes, and error envelopes. |
Systems Integrators, Any Language |
| 06 | 06. Python SDK Reference | typesafe-sdk, Sync & Async Clients, Type Hints, Retry Policy, exception handling. |
Python Developers, ML Engineers |
| 07 | 07. TypeScript / JS SDK Reference | @typesafe-ai/sdk, helper functions (choice, score, noul), Type inference, Promises. |
Frontend / Fullstack Devs, Node.js |
| 08 | 08. Architectural Patterns | 5 Golden Patterns: Speculative Fan-out, Confidence Routing, Composite Scoring, Intent Routing, SDE Cascade. | Software Architects, Leads |
| 09 | 09. Production Cookbooks & Recipes | 10+ end-to-end recipes: Re-ranking, Semantic find, LLM Guardrails, Entity extraction, Tool routing. | Full Development Team |
| 10 | 10. Model Specs & Jagged Edges | Model specs, latency ceilings, token budgets, and the jagged edge catalog to avoid. | Operations & DevOps Engineers |
| Case Study | Empirical Benchmark & Battle Journal | 290-line empirical benchmark report, battle journal, and speculative commit engine architecture. | All Engineers & Decision Makers |
⚡ 30-Second Overview
- What is System One? Unlike autoregressive LLMs designed for generative prose, System One is purpose-built for software: it takes an arbitrary
state+ typedquestionsand returns strongly typed data with calibratedprobabilitiesandconfidencescores. - How does it differ from generative LLMs (GPT-4, Claude, Gemini)?
- Generative LLMs represent System Two: deliberate, streaming text generation, high latency (3–8s), high cost, and fragile JSON schema parsing.
- System One represents Fast Intuition: non-autoregressive, sub-10ms on-device execution (6.53ms with Laya-MLX), 95% cheaper, and zero hallucinations about structure.
- Code in Control: The model serves as programmable semantic common sense embedded within conditional branches (
if/else), while your application code retains deterministic control over state transitions, database mutations, and API calls.