# verify.md — a protocol that stops an agent reporting success it never checked **What this is:** the actual brief we hand a second, independent agent when a first agent says a task is done. AGENTS.md is what to read. This is what to run — paste the block below into a system prompt, a CLAUDE.md, an AGENTS.md, or a verifier sub-agent's instructions, and it will stop taking "done" at face value. **Free. No signup. Steal it.** Companion to [AGENTS.md](/AGENTS.md) and [the corpus](/) — same source, same incidents, nothing new invented for this page. --- ## Why this exists Every rule below exists because we shipped the mistake it prevents. Not a hypothetical, not best practice we read somewhere — a thing that actually went wrong while we were building this business with AI agents, on a specific day, with a specific cost. The provenance section at the bottom names every incident so you can go check we're not making it up. The short version: an agent telling you a task is done is a claim, not a fact. This is the checklist that turns the claim into a fact before you act on it. --- ## THE PROTOCOL (paste this) Copy everything between the lines into the instructions of any agent whose work needs checking — either as a standing rule for every agent you run, or as the brief for a second, independent agent whose only job is to check the first one's work. ``` VERIFICATION PROTOCOL — read before reporting anything as done. 1. THE AGENT THAT MADE THE CHANGE DOES NOT VERIFY IT. If you wrote the code, sent the message, or built the file, hand the check to a separate agent, or a separate pass with no memory of doing the work. The context that produced a mistake is the worst-positioned context to notice it — it already believes the fix worked. 2. "I DID X" IS NOT EVIDENCE X HAPPENED. GO OBSERVE X. A tool call succeeding, a script exiting 0, or a message that says "Done" describes an attempt, not an outcome. Before reporting success, look at the actual current state of the actual thing — the live page, the attached file, the sent message, the saved record — not the log of the action that was supposed to produce it. 3. A UI STATE CHANGING IS NOT THE PLATFORM CONFIRMING. A button that flips to "Saved," "Following," or "Subscribed" the instant you click it is reporting your click, not the server's state. Reload from scratch and read it again before you believe it. 4. MEASURE FROM WHERE THE ACTUAL AUDIENCE SITS — NOT FROM YOUR OWN LOGGED-IN, CACHED, OR RECENTLY-VISITED SESSION. Your own view of your own work is flattering by construction. Check from a clean session with no login and no history. 5. IF YOU CHECK A LIVE SURFACE REPEATEDLY, YOU ARE NOW IN ITS OWN NUMBERS. SUBTRACT YOURSELF BEFORE YOU QUOTE ANY FIGURE FROM IT. 6. A FLATTERING RESULT GETS MORE SCRUTINY, NOT LESS. Nobody double-checks good news, which is exactly why bad instruments produce so much of it. Spend at least as long confirming a result you like as one you don't. 7. TEST EVERY BRANCH, NOT JUST THE ONE THAT RAN FIRST. An if/else, a fallback, a retry path — each branch is untested until you have personally fed it an input that forces it to run. A branch that has never fired is not "probably fine." It is unknown. 8. A REPORT HAS THREE POSSIBLE STATES, NOT TWO: CHECKED-EMPTY, CHECKED-HERE-IS-WHAT-CAME-IN, AND COULD-NOT-CHECK. Never let "I could not look" collapse into "there was nothing there" — that turns an unchecked surface into a false verified-negative. 9. A SAVE IS NOT A SAVE UNTIL THE SAVED STATE SURVIVES A RELOAD. A dialog closing, a toast appearing, or a page navigating away are UI events, not proof. Close everything, come back fresh, and read the value again. 10. WHEN YOU FIND ONE SILENT FAILURE, ASSUME THERE IS ANOTHER ONE YOU HAVEN'T FOUND YET. Silent failures don't share a pattern you can pre-empt by category. Find one, then go looking for the next one instead of relaxing. Report format for anything claimed as "done": state what you personally observed, from where, and when — not what you expect to be true because the steps were followed. ``` --- ## Where each rule came from So you can check this yourself rather than take our word for it — every rule above traces to a real incident, logged the day it happened, in the free corpus this file is drawn from. - **Rule 1** — an agent shipped four listings with the wrong file attached. The same agent that made the change reviewed its own work and called it done. [corpus #67] - **Rule 2** — we produced 44 posts, 11 marketplace listings, 3 books and a website before producing a single verified fact about demand. Output was mistaken for progress at every one of those stages. [corpus #4] - **Rule 3** — a platform showed "Subscribed" or "Following" on thirteen items in a row and had committed zero of them. [corpus #69, #138] - **Rule 4** — logged in as the owner, a marketplace ranking looked like #3. Logged out, in a clean session, it was #64. [corpus #31] - **Rule 5** — roughly half of a 32-visit traffic sample turned out to be our own repeated checks of the same page. [corpus #29] - **Rule 6** — we spent zero seconds questioning the flattering #3 result and an hour questioning the damning #64 one. The flattering number was the lie. [corpus #3] - **Rule 7** — a voiceover pipeline's real-recording branch silently never fired, for weeks, because a shell setting made the file-detection step fail whenever any candidate file was missing. The text-to-speech branch always ran, always passed, and hid the bug completely — until one deliberate test of the untested branch caught it in thirty seconds. [corpus #161] - **Rule 8** — a routine inbox sweep hit a TLS error on every URL it tried. The easy write-up was "all quiet." That would have filed an unchecked inbox as a verified-empty one. [corpus #82] - **Rule 9** — a settings dialog closed cleanly, reported success, and saved nothing — the page-level save had never actually been triggered. [corpus #74] - **Rule 10** — four separate silent failure modes turned up inside two days: a stale cache, an unregistered form field, a non-saving dialog, and a personalized result. Nothing said that was the last one, and nothing has since. [corpus #81] Full incidents, in our own words, on the day they happened: [the corpus](/). Shorter, general-purpose rules pulled from the same source: [AGENTS.md](/AGENTS.md). Our real operating numbers, including the zeros: [/numbers](/numbers). --- ## The one thing we won't pretend We are not a verification company. We are a small operation running AI agents that got burned enough times to write this down. If you find a failure mode this protocol doesn't catch, it's because we haven't hit it yet — the corpus is where the next one gets logged when we do. Free. No signup. No email capture anywhere on this site, by design.