# QWED Protocol — Deterministic LLM Verification

> QWED Protocol is a deterministic trust boundary for LLMs: verify math/logic/code/SQL claims and block unverified outputs before they hit production.

## Install

Save the content below to `.claude/skills/` or append to your `CLAUDE.md`:

## Quick Use

```bash
pip install qwed
qwed init
qwed doctor
qwed test
```

## Intro

QWED Protocol is a deterministic trust boundary for LLMs: verify math/logic/code/SQL claims and block unverified outputs before they hit production.

- **Best for:** agent teams that want deterministic gates (math/logic/code/SQL) before shipping outputs
- **Works with:** Python SDK + CLI; optional Docker; model-agnostic verification boundary (per README)
- **Setup time:** 10–25 minutes

## Practical Notes

- Per README: `pip install qwed` installs core engines (Math, Code, Facts); `qwed[full]` adds SQL + logic/Z3 and more.
- Per README: `qwed test` runs 12 deterministic tests as a pre-production gate.
- Treat verification failures as product signals: log failures, route to fallback, and add tests for recurring mistakes.

## Main

A safe integration pattern:

1. **Wrap your model call** with a verification step (or verify the final answer + key tool outputs).
2. Classify failures into: *correctable* (auto-fix), *requires human review*, and *block*.
3. Start with deterministic domains: math, SQL, code safety checks, logic constraints.
4. Add a CI gate: if your “golden tests” fail, don’t deploy the agent prompt/toolchain update.

Verification works best when it’s explicit: decide what must be true (constraints) instead of hoping the model “usually” gets it right.

### FAQ

**Q: Does verification reduce hallucinations?**
A: Per README: it focuses on making hallucinations irrelevant by catching/blocking unverified outputs before production.

**Q: What’s a good first use case?**
A: Math and SQL checks are low-cost and high-impact. Add code-safety checks for tools that run commands.

**Q: How strict should I be?**
A: Start strict on critical facts/tool actions, but allow fallbacks (re-ask model, different tool, or human review) to avoid dead ends.

## Source & Thanks

> Source: https://github.com/QWED-AI/qwed-verification
> License: Apache-2.0
> GitHub stars: 57 · forks: 9

---

<!-- ZH -->

## 快速使用

```bash
pip install qwed
qwed init
qwed doctor
qwed test
```

## 简介

QWED Protocol 是面向 LLM/Agent 的确定性“信任边界”：内置 SymPy、Z3、AST、SQLGlot 等验证引擎，支持 `pip install qwed` 与 `qwed init/doctor/test`，把不可验证输出在上线前拦住。

- **适合谁：** 希望在上线前用确定性规则验证（数学/逻辑/代码/SQL）的 agent 团队
- **可搭配：** Python SDK + CLI；可选 Docker；模型无关的验证边界（见 README）
- **准备时间：** 10–25 分钟

## 实战建议

- README：`pip install qwed` 安装核心引擎（Math/Code/Facts）；`qwed[full]` 扩展 SQL/逻辑(Z3) 等能力。
- README：`qwed test` 运行 12 条确定性测试，可作为上线前门禁。
- 把验证失败当作产品信号：记录失败样本、做 fallback，并为常见错误补测试用例。

## 主要内容

安全接入方式（从易到难）：

1. **把模型输出包一层验证**（或只验证最终结论 + 关键工具输出）。
2. 把失败分成三类：*可自动修正*、*需要人工复核*、*必须阻断*。
3. 先从确定性强的域开始：数学、SQL、代码安全检查、逻辑约束。
4. 在 CI 里加门禁：黄金测试不过，就不要发布 prompt/toolchain 更新。

验证体系要“显式”：先定义必须成立的约束，而不是指望模型大多数时候都对。

### FAQ

**它会让模型不幻觉吗？**
答：README 的定位是：不是减少幻觉，而是通过验证/阻断让幻觉变得无害。

**最先落地哪个场景？**
答：数学与 SQL 检查成本低、收益高；对能执行命令的工具链再加代码安全检查。

**门禁要多严格？**
答：关键事实与工具动作要严格；但保留 fallback（重问、换工具、人工复核）避免流程卡死。

## 来源与感谢

> Source: https://github.com/QWED-AI/qwed-verification
> License: Apache-2.0
> GitHub stars: 57 · forks: 9


---
Source: https://tokrepo.com/en/workflows/qwed-protocol-deterministic-llm-verification
Author: Agent Toolkit