# design-os-system-one: TypeSafe System 1 & Dual-Brain Architecture

> design-os-system-one: sub-10ms on-device decisions with Laya-MLX, 5-stage speculative commit engine, and 95% LLM cost reduction for autonomous agents.

Sub-10ms Calibrated Judgments for Autonomous Agents with Zero Cloud Waste.

## Documentation & Case Study

- [Landing Overview](https://jangtrinh.github.io/design-os-system-one/): Main project overview, benchmarks, and interactive architecture.
- [Case Study & Battle Journal](https://jangtrinh.github.io/design-os-system-one/case-study.html): In-depth empirical benchmark report, battle journal, and speculative commit engine architecture.
- [Architectural Patterns](https://jangtrinh.github.io/design-os-system-one/08-architectural-patterns.html): Architectural patterns for agents, swarms, and browser loops.
- [Core Primitives](https://jangtrinh.github.io/design-os-system-one/02-core-primitives.html): TypeSafe System 1 decision contracts (Choice, Score, Noul).
- [Full Technical Specification (LLMs Full)](https://jangtrinh.github.io/design-os-system-one/llms-full.txt): Complete architecture specification, empirical benchmark logs, and execution contracts for AI search crawlers.
- [Repository](https://github.com/jangtrinh/design-os-system-one): GitHub source code, releases, and tests.

## Key Architecture & Invariants

- **Dual-Brain System 1 Cascade**:
  - Tier 1: Local Laya-MLX on Apple Silicon Unified Memory (6.53ms P50, $0.00 cost, 100% air-gapped).
  - Cascade sweet spot (tau = 0.30): 70% of traffic resolved on-device in <10ms; 30% ambiguous cases escalate to TypeSafe JEV Cloud API (3.41x overall speedup).
- **5-Stage Speculative Commit Engine**:
  - `OBSERVED -> EVALUATED -> PREPARED -> DISPATCHING -> VERIFYING -> CONFIRMED` (or `UNKNOWN -> RECONCILING`).
  - Speculate on judgments only; high-risk browser mutations require postcondition verifiers and transaction journaling.
  - Dispatch timeouts transition to `UNKNOWN` / `UNCERTAIN` and trigger read-only DOM reconciliation, eliminating duplicate mutations.
- **Revocation of Autonomous 'DONE'**:
  - Models cannot unilaterally close tasks; task completion requires 100% acceptance test proof (`VERIFY_AND_STOP`).
- **Cryptographic exactCache**:
  - All Gateway Router and Skill Pruner cache keys use SHA-256 over canonical JSON to prevent cross-prompt collision exploits.
- **Enterprise Swarm & Penetration-Hardened Security**:
  - `PreExecutionGuardrail`: <2ms pre-dispatch threat firewall intercepting zero-width obfuscation, pipe-to-shell RCE, DAN jailbreaks, and credential leaks (33/33 penetration tests passing).
  - `HierarchicalAgentRouter`: Sub-35ms specialist delegation replacing slow 3,000ms LLM managers with calibrated Softmax and opt-in abstention (`min_confidence: 0.30`).
  - `HighCardinalityShortlist`: Filters 1,000+ candidates down to top-10 in 3.6ms, preventing head token overflow (`head_max_len`).
  - `Browser Reflex Head`: 17–23ms step latency (200x faster than 27B LLMs) for autonomous DOM navigation.
- **Air-Gapped Privacy (`localOnly: true`)**:
  - Strictly forbids WAN cloud calls, guaranteeing zero telemetry and zero external data leaks.