Before AI-written text ships, run the four checks it always fails.
11 files (a small engine + rulesets) · zero dependencies · stdlib only
- The delivery doc still contains '[TBD]' and an 'about 30%' that nobody verified.
- Page 3 says the budget is 80k; page 11 says 120k.
- Placeholders, fake citations, unit drift — a human has to read every page to catch them.
One command, on a file you already have:
python gate.py run ai-output examples/draft.md --format textReal output — produced by the bundled examples/, not a mock-up
==========================================================================
output-gate · 规则集 ai-output v1.1.0
目标：客户方案 v3
场景：AI 产出通用体检
==========================================================================
✗ [PLACEHOLDER](project_custom) 存在未替换的占位符（TBD/待补/XXX/{{}}）—— 发出去就是半成品
→ 2 处占位符未替换：第10行 待补、第11行 TBD
✓ [CODEBLOCK-UNCLOSED](project_custom) 代码块未闭合（``` 数量为奇数）
→ 0 处代码块均闭合
– [BRACKET-UNBALANCED](project_custom) 代码括号不平衡（() [] {} 未配对）—— 直接导致无法运行
→ 非代码文件，不适用
✓ [SECRET-LEAK](project_custom) 疑似密钥/令牌泄漏（sk-…、api_key、token=、private key）
→ 未检出疑似凭据
✓ [PII-LEAK](project_custom) 疑似个人信息（手机号/身份证/银行卡）—— 需人判断是否为公开信息
→ 未检出个人信息特征
⚠ [CLAIM-NO-SOURCE](project_custom) 含统计数字的断言没有出处标记 —— 读者无法验证
→ 2/7 条含数字无出处（29% ≥ 15%）。例：第4行「整体预算约 30 万，其中开发占 60%」；第14行「历史版本预算为 8 万，本版调整为 12 万」
✓ [CLAIM-ABSOLUTE](project_custom) 不可验证的绝对化表述（一定/绝对/100%/永久/零风险）
→ 无绝对化表述
✓ [STRUCTURE-SHALLOW](project_custom) 标题层级过浅（无 h2 及以下）—— 多半是平铺直叙，难读
→ 5 个标题，4 个二级及以下
⚠ [SIZE-ANOMALY](project_custom) 篇幅异常（过短疑似敷衍 / 过长疑似灌水）
→ 仅 16 行，低于 20 行 —— 疑似敷衍
✓ [COMPLETION-NO-EXIT](project_custom) 自报完成（已完成/跑通/全部通过…）但全文无任何退路声明（未通过怎么办/回滚/降级/视为未完成）——按 ZEG 降级为「未验证的自报」而非完成
→ 无自报完成表述（ZEG 不适用）
--------------------------------------------------------------------------
合计 10 条：BLOCK 1 ｜ WARN 2 ｜ NEEDS_VALIDATION 0 ｜ INFO 0 ｜ PASS 6 ｜ SKIP 1
⛔ 本规则集做【形式与可追溯性】检查，不判断内容对错
⛔ WARN 项需人判断 —— 不要把 WARN 当成「有问题」的证据
⛔ 「无源数字」不等于「数据是假的」，只等于「读者无法验证」
结论：不通过（存在 BLOCK 项），先修再发布。
- Rule-driven checks: placeholders, unsourced numbers, numeric conflicts, unit drift.
- Emits text / json / sarif / junit — SARIF lands straight in code scanning.
- Keeps a coverage ledger across runs: what newly failed, what you actually fixed.
- It does not verify facts. It only checks that numbers declare a source.
- Rulesets are opinionated and domain-bound (ai-output, design, ecommerce, music, shortvideo).
- It will not catch a plausible lie written fluently.
Being explicit here is cheaper than a wrong expectation later.
52 rules across five rulesets — one engine, one directory per vertical.
python gate.py list prints this list live, with each ruleset's own version.
| Ruleset | What it checks | Rules | Version | Run it (exits 1) |
|---|---|---|---|---|
ai-output |
AI 产出通用体检 | 10 | v1.1.0 | python gate.py run ai-output examples/draft.md |
design |
设计输出体检 | 5 | v0.1.0 | python gate.py run design rulesets/design/fixtures/design_fail.html |
ecommerce |
跨境电商 listing 合规 | 10 | v2026-10-11.0 | python gate.py run ecommerce rulesets/ecommerce/fixtures/listing_fail.json |
music |
音乐发行前合规体检 | 20 | v2026-10-04.1 | python gate.py run music examples/track_fail.json |
shortvideo |
短视频 / 口播脚本合规 | 7 | v1.0.0 | python gate.py run shortvideo rulesets/shortvideo/fixtures/script_fail.json |
The bundled examples/ and rulesets/*/fixtures/ are committed, so every
command above reproduces on a fresh clone. The clean fixture of each ruleset
exits 0 and the failing one exits 1; CI asserts both directions, so a green
badge means the ruleset fired when it should and stayed quiet when it should.
Adding a vertical means dropping in a directory — the engine does not change.
0 = clean (no BLOCK, no WARN). 1 = at least one BLOCK rule fired. 2 = WARN-level findings only — nothing blocks the build. 3 = you asked for something this gate could not do: a malformed command line, an unknown ruleset, an unreadable file, or an engine failure (it prints ✗ … first). 2 used to double as the usage-error code; since 2026-10-20 a mistyped command exits 3, so a non-zero badge always means a real finding.
git clone https://github.com/chenhz01/output-gate
cd output-gate
python gate.py --helpPython 3.9+. Nothing to pip install — stdlib only.
- name: output-gate
run: python gate.py run ai-output examples/draft.md --format textaction.yml is included, so you can call this repo directly as a composite action.
# .github/workflows/gates.yml
- uses: actions/checkout@v4
- uses: chenhz01/output-gate@v0.1.3
with:
# IMPORTANT: this path is inside YOUR repository.
# The examples/ in this repo are NOT in your workspace — pass your own files.
args: "run ai-output path/to/draft.md --format text"Pin the tag (@v0.1.3) rather than @main — main moves.
Check Releases for the newest tag; v0.1.0 predates the
fix that makes a mistyped path fail instead of silently passing.
If you omit args, the action only prints help and exits 0. It raises a warning
annotation — quoting this gate's own correct usage — so a green check is never
mistaken for "it inspected my files".
产出体检门 — AI 写的东西发出去之前，先跑一遍它最常犯的四类错。
- 交付文档里留着「[待补充]」和一个没人核过的「约 30%」。
- 同一份方案里，第 3 页说预算 8 万，第 11 页说 12 万。
- 占位符、伪引用、单位错乱——每一份都要人肉读一遍才发现。
一行命令：python gate.py run ai-output examples/draft.md --format text
边界：It does not verify facts. It only checks that numbers declare a source.
十五个链接是迷宫，一个症状是门。按上表挑一个，其余九个在下面 The gates。
每个仓都能单独活：复制 gate.py，一条命令跑完，不共享运行时、不装包、不加配置。
有星就给这一个，没有也不影响任何仓的可用性——没有任何门是「先给星才能用」。
本页承诺的是 15 项可机器复核的结构特征，当前 15/15（2026-10-20 复核）。
GitHub 从未公开 Trending 的排序算法，所以任何「保证上热门」的说法都在描述一件它不可能知道的事；
本仓不这么说。跑 python tools/hotness.py 可自行复算每一条。
Fifteen links is a maze; one symptom is a door. Find your symptom, install that gate, ignore the rest until it hurts.
| what you are seeing | install this |
|---|---|
| the bill is 10x normal and the trace is a wall of the same step | loop-gate |
| the agent can delete prod and nobody signed off on that | permission-gate |
| answers get worse after truncation and nobody knows what got dropped | context-gate |
| two rules in the prompt contradict each other; the model obeys neither | instruction-gate |
| AI text ships with fabricated links and empty phrases | output-gate ← you are here |
| half the repo might be dead and nobody can prove which half | deadcode-gate |
A gate you did not need is a gate you uninstall, so nothing here says "install all
fifteen". Each repo stands alone: copy gate.py, run one command, no shared runtime,
no package manager, no config file to grow.
On stars. If this one saved you an incident, star it — that is the only signal I get about which gate to build next. There is no exchange: starring one repo does not owe me a star on another, and no gate here is unlocked by a star. Star trading breaks GitHub's rules and would poison the single honest metric this project has.
What is guaranteed is structure, not ranking. Fifteen properties, each one machine-checked against this repo — current score 15/15 (2026-10-20).
Plain version: instead of asking you to trust a badge, I list what I actually checked, name the script that checks it, and let you re-run it. If one of these lines is false, that is a bug in my repo, not a matter of taste.
- Falsifiable side:
python tools/hotness.pyre-runs every check and prints the evidence line for each one. You do not have to trust this paragraph. - Non-falsifiable side: GitHub has never published how Trending sorts repositories, so any repo promising you a guaranteed spot is telling you something it cannot know. This one does not make that promise.
- If a claim on this page stops being true, that is a bug: open an issue and it gets fixed in the same week, not in the next launch.
What actually carries this repo:
- the command in this README is byte-for-byte the command CI runs (53-char argv, exact match)
- the output block is a real run of
examples/, not a mock-up (Real output section) - CI asserts the exit code, so the badge can turn red — and it has (asserts exit=1)
- every fixture is committed, so a fresh clone reproduces the output (1 file under examples/ tracked in git)
* -textin.gitattributes: same bytes on Windows and Linux (.gitattributes * -text)
| # | gate | what it stops |
|---|---|---|
| 02 | loop-gate | Find where your agent looped, before it burns another run |
| 03 | permission-gate | Score how much more your agent can do than it should |
| 04 | context-gate | Measure what truncation deletes, before the model starts guessing |
| 05 | instruction-gate | Two of your instructions contradict each other. This finds them |
| 06 | output-gate ← you are here | Before AI-written text ships, run the four checks it always fails |
| 07 | memory-expiry-gate | Give every memory a TTL, then prove the expiry semantics actually hold |
| 08 | drift-gate | When embeddings drift, retrieval quietly rots. Fingerprint it |
| 09 | price-gate | Your cost model is only as good as the price table under it |
| 10 | complexity-ceiling | Per-file complexity budgets that fail the build instead of the reviewer |
| 11 | deadcode-gate | Quantify how much of the repo is dead, instead of guessing |
| 12 | docrot-gate | Docs don't die at once. They rot one stale command at a time |
| 13 | leak-gate | Scan for internal detail before the doc leaves the building |
| 14 | backup-gate | A backup you have never verified is a rumor |
| 15 | trail-gate | Hash-chained agent audit trail — and it verifies its own chain |
- Code: Apache-2.0 — see LICENSE
- Docs and example data: CC BY 4.0 — see LICENSE-SPEC
Attribution required when you redistribute the documentation; code can be used commercially without a separate agreement. See NOTICE.