Audison

KO | EN

gitlite — search

#react #llm #go #machine-learning #deep-learning #linux #macos #java #nodejs #windows #awesome-list #vue #security #mcp

(★ 39)

Stop trusting single AI outputs. An anti-hallucination framework with dual-brain arbitration, multi-model audit, and evidence chains for LLM quality control.

Audison Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Two AIs argue. One verdict.

---

**Real numbers, not benchmarks.** Audison audited [httpx v0.28.1](https://github.com/encode/httpx) (14,000+ Stars) across 10 security-critical functions using **EmpiricalVerifier** (AST static analysis — a utility distinct from Audison's multi-model adversarial audit).

| CONFIRMED | REFUTED | False Positive |
|:---------:|:-------:|:--------------:|
| 4 | 4 | 1¹ |

> ¹ `_get_client_nonce` was an AST false positive (indirect variable assignment untraceable), excluded after manual review. One UNCERTAIN excluded: `_build_auth_header` same-name ambiguity, a tool limitation, not code uncertainty. [Full report →](reports/real_world_audit.md)

```bash
pip install audison && audison scan .
```

<details>
<summary>Example output: scanning Audison itself → 0 CONFIRMED, 43 REFUTED [full output →](docs/terminal_output.txt)</summary>

```
 Audison Scan — 维度4 数据完整性实证验证
 扫描目录 D:\HANAKO\audison
 发现文件 104 个源文件 / 审查函数 83 个
 CONFIRMED 0 (真阳性)
 REFUTED 43 (假阳性)
 UNCERTAIN 40 (需人工复核)
```
</details>

---

## The Problem

A single model can't discover its own blind spots. Same training data, same biases — one person grading their own homework. Audison makes two models argue: one audits, another attacks from 5 adversary perspectives, a third cross-validates. Consensus comes from **surviving attack**, not from agreeing to agree.

---

## What Audison Catches That Others Miss

Same AI-generated auth code, four reviewers:

| Finding | Standard AI Review | CodeQL | Shannon | Audison |
|---|---|---|---|---|
| SQL injection in login | Missed | Missed | Missed | Found |
| Hardcoded JWT secret | Warning | Missed | Missed | Found |
| Missing rate limiting | Missed | Missed | Missed | Found |
| CSRF token bypass | Missed | Missed | Missed | Found |

---

## How It Works

```
Input Code
    │
    ▼
[ Brain One ]  ──── Primary audit (security, correctness, logic)
  (GPT-4o)
    │
    ▼
[ Opponent Brain ] ──── 5 adversarial perspectives attack findings
  (Claude Sonnet)
    │
    ▼
[ Brain Two ]  ──── Cross-verification: consensus → confirmed,
                     disagreement → UNCERTAIN (not hidden)
    │
    ▼
[ TrustReport ]  ──── Verdict + Confidence + Findings + SHA-256 chain
```

---

## Quick Start

```bash
pip install audison
audison scan .              # EmpiricalVerifier: AST static analysis
audison audit .             # Full multi-model adversarial audit

# Set API keys (two providers recommended for cross-arbitration):
export OPENAI_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."
```

---

## Comparison

| Feature | Audison | Standard AI Review | CodeQL | Shannon |
|---------|:-------:|:------------------:|:------:|:-------:|
| Open Source | Yes | — | Yes | No |
| Multi-model Arbitration | Yes | No | No | Yes |
| Adversarial Review | Yes | No | No | No |
| Uncertainty Transparency | Yes | No | No | No |
| Verifiable Evidence Chain | Yes | No | No | No |
| Cost | Free; your API keys | Free | Free | Subscription |

---

## Project Structure

```
audison/
├── src/audison/
│   ├── engine/              # TrustEngine — standalone audit layer
│   ├── brains/              # Brain One, Brain Two, Opponent Brain
│   ├── core/                # Caching, context, session management
│   └── utils/               # LLM client (8 providers), token counter
├── tests/unit/              # 186 unit tests
├── reports/                 # Audit reports
├── docs/                    # API, integrations, getting started
├── pyproject.toml
└── LICENSE
```

---

## License

---

AI proposes. AI challenges. You decide.

<a href="https://wdnmd1265.github.io/Audison/playground.html">Try it live: playground.html</a>
 &nbsp;|&nbsp;
 <a href="https://wdnmd1265.github.io/Audison/">GitHub Pages</a>

---

**详细文档见 [docs/](docs/)** — API 详解、MCP Server 配置、GitHub Action、LangChain/CrewAI 集成等。

Audison

Audison Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Audison

Audison Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README