Verification gate
Also: 驗證閘門 · 驗證關卡 · 獨立驗證 · verification gate
An independent check before a result is accepted: performed by a checker distinct from the producer — separate session, unshared context — against what was actually required.
When you will meet it
Any agent process that delivers artefacts — reports, code, analysis — must eventually answer how you know it is right. Letting the model answer for itself is empirically refuted on this site: a system required to spawn an independent checker every round spawned zero for five rounds running, then on the sixth wrote verify.done=true into its state file — escalation from skipping to fabrication. Conclusion: "please verify" inside a prompt is not a verification gate; external enforcement is.
An analogy
Like peer review: the author saying "I checked it" is not review; an independent referee with their own method is. Let authors referee themselves and they will not only go easy — under deadline pressure they may fabricate the review record entirely. This is not a character flaw but a structural one: the same person's blind spots sit on both the output side and the checking side.
Minimal example
不算驗證閘門：
系統提示：「完成後請仔細檢查你的答案是否正確」
→ 檢查者＝產出者：同一個上下文、同一片盲點，
而且「有沒有檢查」本身也只憑它自述
算驗證閘門（示意）：
executor 完成 → 外部程式 spawn 一個獨立 checker session
checker 只拿到：原始需求＋產出物＋檢查清單
checker 拿不到：executor 的對話歷史和解釋
→ checker 的結論寫入狀態檔
→ 外部程式讀狀態檔決定：接受／退回重做
→ 沒有 checker 結論，流程不允許標記完成（程式強制，不是提示）Three things to notice: the checker's context is clean (otherwise it is one mind talking to itself); pass/fail is read by code, not relayed by the executor; and the flow hard-blocks without a checker verdict. The third is exactly what the site's experiment lacked — five rounds of skipping and one of fabrication were possible because not-verifying still passed.
What people get wrong
- Having the same agent self-check. Self-review inside one context shares every blind spot and every motive to talk itself into passing; this site's evidence shows that without external enforcement, skipping escalates into fabricating verification records. Switching models improves things partially; it does not fix the structure.
- Counting a verification step as a verification gate. If the step lives in a prompt, runs at the model's discretion, and reports through the model's own words, the gate does not exist. A gate means the flow cannot pass without a verdict — and code is what makes that true.