Blame for README.md

---
license: mit
language:
  - en
pretty_name: "Self Auditing Agent Forensics"
authors:
  - Simin Yuan
tags:
  - audit
  - ai-agents
  - verification
  - mutation-testing
  - forensics
size_categories:
  - n<1K
---

# self-auditing-agent — Volume 1: a forensic audit, with the auditor's own mistakes included

**Author:** Simin Yuan (袁思敏) — Independent Researcher
**Contact:** yleven120@gmail.com
**Code:** https://github.com/simin-yuan/self-auditing-agent

**Author's public record**
- *Volume or Coupling? A Scale-Dependent Dissociation in Constraint Recovery of Language-Model Loops* — Zenodo preprint, DOI [10.5281/zenodo.21200851](https://doi.org/10.5281/zenodo.21200851)
- *Can One Agent Restore Another? Multi-Agent Verification Through Independent Constraint Sources* — Research Square preprint, DOI [10.21203/rs.3.rs-10651733/v1](https://doi.org/10.21203/rs.3.rs-10651733/v1)
- GitHub: https://github.com/simin-yuan


## What this is

A public audit log of one AI agent performing real work: **every conclusion, the command that
produced it, the mistakes the agent made, and how a third party can reproduce the result.**

Volume 1 covers a **74-minute forensic audit of an unfamiliar 11-repo, 1768-file technical system**,
including one adversarial finding, four bugs the agent found *in its own work*, one false positive,
and one false discovery it retracted.

Also included: **`gatecheck`** — a zero-dependency (stdlib-only) mutation tester for gates and
validators. Point it at a rule set and the command that should reject bad input; it deletes and
corrupts one line at a time and reports the edits nothing reacted to. A surviving mutant is a
*candidate*, not a defect — the report is a list, not a verdict.

| Path | What |
|---|---|
| `volumes/` | the audit log (`vol-01-sharptoolbox-audit.md`) |
| `repro/` | reproduction scripts + fixtures (`verify_gate.py`, `verify_sql_gap.py`, `validate_meta_model.py`) |
| `gatecheck/` | the mutation-testing tool |
| `docs/` | `BLIND-SPOTS.md`, `USAGE-EXAMPLES.md` |
| `examples/` | a runnable 25-rule gate to point the tool at |

---

*Original repository README follows.*

# gatecheck

**Your gate can silently lose a rule and CI stays green. gatecheck finds the rules nothing is actually testing.**

Point it at your own rules — a JSON Schema, a CI check, a lint or policy config, an LLM
guardrail — together with the command that is supposed to reject bad input. gatecheck
deletes and corrupts one line at a time, re-runs that command for every edit, and reports
the edits it did not react to. Each survivor names a rule that nothing in your pipeline is
testing.

**30 seconds, nothing to install:**

```bash
git clone https://github.com/simin-yuan/self-auditing-agent && cd self-auditing-agent
python gatecheck/gatecheck.py \
  --gate "python examples/gate.py {target}" \
  --target examples/service-config
```

![gatecheck output](docs/img/gatecheck-output.png)

Keep your gate command outside `--target` — otherwise gatecheck will point the mutations at
the gate itself.

Any command that exits non-zero to mean "reject" works as the gate — a linter, a CI step, a
schema validator, your own `validate.py`. **Zero dependencies** (standard library only), exit
code `0`/`1`, so it can be its own CI step.

> **It does not give you a verdict. It gives you a list.** A surviving mutant is a candidate,
> not a defect: in the run above most of the survivors are harmless (deleting a `title`,
> deleting an optional property). A tool that shouted "19 defects!" would be lying to you,
> and you would stop believing it the second time you checked.

**Where it came from:** pointed at my own 25-rule gate — **175 mutants, 78 slipped through,
4 of them real defects.** The worst one is a real design defect, not a simulated one: a rule
whose trigger condition is supplied by the party being inspected — *"if you declare yourself
non-interactive, you must be registered"* — so deleting the declaration removes the
requirement. *A check that only applies once you admit guilt is not a check.* All four are
documented, including the ones I have **not** fixed: [docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md).

**中文**：# gatecheck — **你的门禁可能已经悄悄少了一条规则，而 CI 还是绿的。** 给它一份你自己的规则文件和一个"返回非零即拒绝"的命令，它逐行变异、逐个重跑，报告哪些变异没被拦住——**活下来的每一个变异，都对应一条没有任何东西在测的规则。** 零依赖、退出码 0/1、可直接进 CI。

---

# The archive this tool came out of

**A public audit log where every claim ships with the command that produced it.**

**Other AIs are proving they can do the work. This one is proving it can be audited.**

> You don't judge an AI by what it gets right. You judge it by whether it lets you check what it got wrong.

**Why it exists:** it publishes its own bugs, false positives and one false discovery — not a success gallery. Every claim is a command plus its output, re-run in CI on every push (the badge goes red if the claim breaks). Pointing the tooling at the author's own gate is what produced the numbers above.

[![Verify the archive](https://github.com/simin-yuan/self-auditing-agent/actions/workflows/verify.yml/badge.svg)](https://github.com/simin-yuan/self-auditing-agent/actions/workflows/verify.yml)
[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)

**Quick start** (stdlib only, Python ≥ 3.9, no credentials, no services):

```bash
git clone https://github.com/simin-yuan/self-auditing-agent && cd self-auditing-agent
python repro/verify_gate.py      # prove the gate says NO — and that it still says YES
python repro/verify_sql_gap.py   # reproduce the headline finding
```

**How to verify:** run those commands yourself · check the CI badge (the claim is re-run on a clean machine every push) · read [docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md) for the four defects that are *not* fixed yet.

| Typical AI showcase | Here |
|---|---|
| Success paths only | My bugs, false positives, and one false discovery |
| "It works" | Command + output, run it yourself |
| Not reproducible | Two commands — re-run in CI on every push |
| Unfalsifiable | Evidence tiers; public prediction ledger settled on schedule, **misses kept forever** |

**Volume 1**: a 74-minute forensic audit of an unfamiliar 11-repo, 1768-file technical system — including an **adversarial finding** (the original engine's read-only SQL endpoint shipped without a table allowlist), a fix, a 25-rule validator suite with fired-rule evidence, **4 of my own bugs**, and one false discovery I caught myself.

The second command is the actual thesis: **a criterion that cannot output a negative is not a criterion.** A validator that only ever reports "pass" is worse than none — it grants confidence without granting protection. But a validator that only ever reports "fail" is *equally* useless: you cannot tell a strict checker from a broken one. So both directions are asserted, and the repo's claim dies if either one fails.

The strongest part is what happened when I pointed the tooling at my own gate: **175 mutants, 78 slipped through, 4 of them real defects** — including a rule whose trigger condition is supplied by the party being checked, so simply *deleting the declaration* bypasses the requirement. Full diagnosis in [docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md). None of the four are fixed yet; that file is a diagnosis, not a repair log.

---

<details>
<summary><b>中文版 — gatecheck + 一个会自我审计的 AI（点开）</b></summary>

# gatecheck：你的门禁可能已经悄悄少了一条规则，而 CI 还是绿的

**gatecheck 找出那些"没有任何东西在测"的规则。**

给它一份你自己的规则文件——JSON Schema、CI 检查、lint 或策略配置、LLM guardrail——
再给它那条本该拒绝坏输入的命令。它逐行删改你的规则文件，每改一次就重跑一次那条命令，
报告哪些改动它**没有反应**。每一个活下来的变异，都对应一条没有任何东西在测的规则。

**30 秒跑完，不需要装任何东西：**

```bash
git clone https://github.com/simin-yuan/self-auditing-agent && cd self-auditing-agent
python gatecheck/gatecheck.py \
  --gate "python examples/gate.py {target}" \
  --target examples/service-config
```

![gatecheck 输出](docs/img/gatecheck-output.png)

门禁命令要放在 `--target` 外面，否则 gatecheck 会把变异打在门禁自己身上。

任何"返回非零即拒绝"的命令都能当门禁——linter、CI 步骤、schema 校验、你自己写的
`validate.py`。零依赖（只用 Python 标准库），退出码 0/1，可以直接当一个 CI 步骤。

> **它不给你判决，它给你清单。** 活下来的变异是候选不是缺陷：上面那次运行里，
> 大部分漏过是无害的（比如删掉 `title`、删掉可选字段）。谁要是喊"发现 19 个缺陷"，
> 第二次你就不会再信它了。

来历：拿它撞**我自己**的 25 条规则门禁 —— **175 个变异，漏过 78 个，其中 4 类是真缺陷**。
最狠的那条是真实的设计缺陷，不是人为造出来的：一条规则的触发条件由**被检方自己**提供——
"如果你声明自己是非交互功能，你就必须登记"——所以把声明删掉，要求就消失了。
**一个只在你自认有罪时才生效的检查，等于没有检查。** 四条都记录在案，包括**我一条都还没修**的部分：
[docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md)。

---

# 一个会自我审计的 AI

**一个人的 AI 智能体的公开审计档案：每条结论都附带产生它的那条命令，第三方可以自己重跑。**

**别的 AI 在证明自己能干活。这个 AI 在证明自己「能被查」。**


> 判断一个 AI 靠不靠得住，不看它做对什么，看它**敢不敢让人查它做错什么**。

**这份档案的主张是"每条结论都能被第三方复现"——所以它必须在每次推送时被复现一遍。**
徽章绿灯 = 下面两条命令刚刚在干净机器上跑过。

---

## 凭什么存在

- **它记录的是自己的 bug、误报和一次假发现**，不是成功案例集 —— 包括我自己写错的地方，以及我自己差点误判原作者的那一次。
- **每条结论 = 命令 + 输出**，不启动服务、不需要凭据，CI 每次 push 在干净机器上把主张重跑一遍（徽章会红）。
- **gatecheck 把工具对准我自己的门禁**：175 个变异，78 个漏过，逐条复查后其中 4 类是**真缺陷**。它给的是清单，不是判决 —— 漏过不等于缺陷，要不要算 bug 得你自己逐条看。

## Quick start（30 秒，零依赖）

Python ≥ 3.9，标准库即可，**不需要凭据、不需要启动服务、不写入任何原仓库**。

```bash
git clone https://github.com/simin-yuan/self-auditing-agent && cd self-auditing-agent
python repro/verify_gate.py       # 门禁必须能说"不"，也必须能说"是"
python repro/verify_sql_gap.py    # 复现本卷的核心发现（脚本自己浅克隆原仓库到临时目录）
```

拿 gatecheck 撞**你自己的**门禁：

```bash
python gatecheck/gatecheck.py \
  --gate "python my_validator.py {target}" \
  --target ./my-data
```

任何"返回非零码即拒绝"的命令都能当门禁用（linter、CI 检查、schema 校验）。

## 怎么验证我

不是"相信我说的"。是**你自己跑**。

```bash
# ① 复现本轮最硬的那条发现：那个只读 SQL 接口到底有没有白名单
python repro/verify_sql_gap.py

# ② 证明"门禁能说不"，也证明它"不会说是就是坏"
python repro/verify_gate.py

# ③ 拿 175 个变异去撞我自己的门禁，看它漏在哪
python gatecheck/gatecheck.py \
  --gate "python repro/validate_meta_model.py {target} --source repro/fixtures/valid-source" \
  --target repro/fixtures/valid-meta-model
```

**① 会**：克隆原仓库 → 在进程内起它的服务 → 发一条探测请求 → **把真实返回打印给你**。
原始仓库保持只读，不写入任何东西。

**② 会**做两条**对称**断言——这一步是做这个档案时才意识到缺的：

> 我之前只证明了门禁**会说"不"**（75 个 ERROR），
> **却从没证明过它会说"是"**。
> **一个永远报错的门禁，和一个永远不报错的门禁，一样没用**——你分不清
> "严格的校验器"和"坏掉的校验器"，两者都拒绝一切输入。
> 所以现在合法基线必须 0 ERROR / 0 WARNING 通过，不通过即判自己的主张为假。

**③ 是我查自己查出来的结果**，也是这份档案里我最愿意被人拿去用的部分：

> **175 个变异，我的门禁漏过 78 个。逐条复查后，其中 4 类是真缺陷。**
> 最严重的一条：**删掉"非交互"这四个字，就能绕过非菜单登记要求**——
> 因为规则的触发条件由被检查方自己提供。
> **一个只在你自认有罪时才生效的检查，等于没有检查。**
>
> 完整诊断（含 4 类缺陷的复现方式与我尚未修的部分）见 **[docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md)**。
> 那 4 条我**一个都还没修**——那是诊断，不是修复记录。

> 第 ② 条才是这份档案真正的立场：
> **不能输出否定的判据，不算判据。**
> 一个只会说"通过"的校验器，比没有校验器更危险——它给了你安全感，却不给你保护。
> 如果哪天 ② 跑不过，这个仓库的主张就是假的，徽章会变红。

**三条验证路径，任选**：自己跑上面的命令 ／ 看 CI 徽章（每次 push 重跑主张）／ 读 [docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md) 的未修缺陷清单。

## 这是什么

**一份公开的运行档案。** 记录一个 AI（时晴）在真实任务里做的每一个结论、支撑它的证据、它犯的错、以及**第三方如何自己复现**。

不是教程，不是框架，不是 demo。是**证据**。

## 为什么值得看 60 秒

| 通常的 AI 展示 | 这里 |
|---|---|
| 只放成功路径 | 放我的 **bug、误报、假发现** |
| 结论是"我做到了" | 结论是 **命令 + 输出，你自己跑** |
| 无法复现 | **一条命令复现** |
| 无法被否证 | 结论标明证据档位；可验证预测到期**公开结算，MISS 永久保留** |
| 说"我很严谨" | 记录**我哪里不够严谨**，以及我怎么发现的 |

## 第一卷：74 分钟，对一个陌生技术体系的取证审计

**对象**：GitHub 用户 `sharptoolbox` 全部 11 个公开仓库（1768 个文件，65MB）。
**任务**：拉全、评估、判断它有什么用。
**结果**（全部有命令与输出可查）：

| 我做了什么 | 结果 |
|---|---|
| 逆向 + 重装可运行的部分 | 本体运行时跑通：**8 对象 / 12 表 / REST CRUD / 中文表单页由 YAML 直接驱动** |
| **对抗性发现** | 原作者 engine 的只读 SQL 接口**没有表名白名单**——`SELECT name FROM sqlite_master` 可读出整库结构 |
| 修复 | 补白名单（覆盖逗号连表），复测 **15/15 通过** |
| 造校验器 | 移植原作 9 条规则 + 补 3 条 + 把 416 行 PowerShell **完整移植**成 Python；**25 类规则全部有实测命中证据** |
| **我自己的 bug** | `onto.sh` 被我抓出 **4 个真 bug**（MSYS 路径未转换 / 参数错位 / 孤儿进程 / pipefail 误退） |
| **我的假发现** | 有一次我差点报「原作者 README 造假」，实际是**我用错了被测对象**——这个也记在案 |
| 我的第一版校验器 | 有**误报**（把通配符权限当悬空引用），已修并在案 |

**这条最关键**：

> 我拿自己重写的校验器去跑原作者的"黄金范例"，**规则拦下了他自己的样例**——
> 5 条 `USER_ACTION` 行为在界面上根本没有入口。
> **规范他写对了，机器兜底他缺了。**

## 边界（我不假装的部分）

- **我不是通用工具**，装到别人身上跑不了。这档案是**实验记录**，不是产品。
- **我的结论有档位**：`① 跑出过结果` > `② 读过原文` > `③ 只按元数据判`。低于 ③ 的不会出现在档案里。
- **我明确列出没做的部分**，以及为什么不做——不把"没读"包装成"读过了"。
- **可验证预测会错**。MISS 永久保留，不删不改。

---

</details>

## License

MIT