> For the complete documentation index, see [llms.txt](https://dsg-3.gitbook.io/dsg-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://dsg-3.gitbook.io/dsg-docs/control-plane/dsg-architecture-deep-dive.md).

# DSG Architecture Deep Dive

## Overview

DSG (Deterministic Execution & Governance) is a runtime governance engine that makes AI decisions provable and auditable.

```
Client Request
    ↓
Policy Parser (Thai/English support)
    ↓
Constraint Normalizer
    ↓
    ├─ QUBO Builder (for optimization problems)
    ├─ Z3 Encoder (for logic/constraint problems)
    └─ Hybrid Router (complexity-based selection)
    ↓
    ├─ Z3 SMT Solver (formal verification)
    └─ QUBO Optimizer (NVIDIA Ising + fallback)
    ↓
DSG Policy Gate
  - Compares deterministic result with optional LLM advisory
  - Deterministic verdict ALWAYS decides (LLM never overrides)
  - Disagreement flags for human review
    ↓
Deterministic Execution
  - Apply solver result
  - Execute agent/workflow with proven context
    ↓
Evidence Store (Supabase)
  - SHA-256 hashed proof chain
  - CCVS L1-L5 compliance artifacts
  - Auditable 2+ years back
```

## Core Layers

### 1. Policy Parser

Converts natural language policies to structured constraints

### 2. Constraint Normalizer

Normalizes constraints to QUBO or Z3 format

### 3. Solver Router

Routes problems to Z3 or QUBO based on complexity:

* Small problems (complexity < 50): Use Z3 (faster)
* Large optimization (complexity > 50): Use QUBO + Z3 verification

### 4. Z3 Formal Verification

Z3 SMT Solver verifies all solutions against constraints

### 5. QUBO Optimizer (NVIDIA Ising)

QUBO optimization for large combinatorial problems

### 6. LLM Advisory Verifier (Optional)

NVIDIA `ising-calibration-1-35b-a3b` provides secondary analysis

* Never modifies solution
* Deterministic verdict always wins
* Disagreement flags for human review

### 7. DSG Policy Gate

Combines deterministic result with optional LLM advisory

### 8. Evidence Store

Records proof chain with SHA-256 hashing for auditability

## Performance

| Operation            | Time           |
| -------------------- | -------------- |
| Policy Parse         | 10-50ms        |
| Constraint Normalize | 20-100ms       |
| Z3 Solve (small)     | 100-500ms      |
| Z3 Solve (medium)    | 500ms-2s       |
| QUBO Optimize        | 100-1000ms     |
| Evidence Record      | 10-50ms        |
| **Total Pipeline**   | **300-1500ms** |

## Determinism Guarantee

Same input + seed → Same output always

```
quboHash + solverVersion + solverSeed + normalizationVersion
→ Identical proof hash
→ Reproducible decision
```

See VERIFICATION.md for details.