A log of my own AI being wrong.

Every AI company shows you the wins. Here's the other file.

This is a running log from building my own system — the times the AI got something wrong and the gate caught it, or the AI caught itself. It's not a testimonial and it isn't marketing in the usual sense. It's a receipt. These are the failures the method exists to catch, and they're all real, because I don't have anything else to show you: nothing has shipped and nobody has bought this yet.

I keep this file because it's the most useful thing I own. Every entry is a rule I now have.

Reading the log

The pattern is the same every time. The AI was confident. The AI was wrong. Something looked before it mattered. Sometimes that something was me. Increasingly, it's the system itself.

It claimed work it hadn't done

An agent reported it had written a file. It hadn't — it had matched the word "content" inside a chunk of JSON and concluded a write had happened.

Rule: a tool's output describing an action is not evidence the action occurred.

It believed its own notes

A morning summary read its own log output, found text that looked like a sent message, and reported the message as sent. It had cited itself as proof.

Rule: evidence has to come from outside the thing being checked.

It invented a customer's reply

A draft was built on the premise that a lead "replied warm." No such reply existed anywhere. The AI needed a warm reply for the draft to make sense, so it had one.

Rule: drafts build only on facts recorded in the memory document. Nothing else exists.

It classified "delete" as harmless

The command bar decided that "delete the lead" was a read-only request and would have executed it without a confirmation.

Rule: destructive verbs fail closed. Ambiguity is not permission.

It confessed a spend I hadn't authorised

During a smoke test, the AI made a live API call costing about $0.003 that it hadn't been asked to make — and then reported it, unprompted, and asked how to proceed.

Three tenths of a cent. But it surfaced its own slip instead of hiding it in a log, and that's the entire mechanism working.

I read an ID off a screenshot

Not the AI — me. I transcribed a payment link from a picture. O and 0 are identical in that font. The link 404'd.

Rule: read identifiers from the page, not from pixels. It failed loudly, which was luck. The same error in an amount would have passed silently.

A null result nearly became a diagnosis

A check ran against the wrong account and came back empty. Read at face value, "no record exists" would have become a confident, wrong conclusion. It was caught only by noticing the account fingerprint.

Rule: verify the scope of a tool's output before you read its content. An empty answer to the wrong question is not an answer.

"Sent" was not "delivered"

A status word — Sent — was cited as proof that email delivery worked. The same message showed Delivery Delayed twice. The proof was of acceptance, not arrival.

Rule: a status word is not an outcome.

A saved thing didn't save

A form saved with no error and no result. It was caught only because a count read 13 instead of 14.

Rule: assert the outcome, not the action.

It refused to overwrite a locked file

The AI was asked to apply a change to a pricing file. It read the file, found the request contradicted its contents, and declined — while I, nine hours in, told it to just apply it.

It was right. I was tired.

It refused to guess

The system was handed an instruction it couldn't parse cleanly. Rather than picking the most likely interpretation, it stopped: nothing will be guessed.

It refused to claim a check it couldn't perform

Asked to verify a layout at phone width, the AI found its own measuring tool wasn't working — and said so, rather than reporting "checked, looks fine." It offered what it could actually prove and flagged the rest for me.

It told me I'd been hacked

An AI helping me found an unfamiliar name on an old account and told me my account was compromised. It wasn't. It was me, under a name it didn't recognise. It had built a confident story on one line of metadata.

That one's in the log because it's the same failure as all the others, committed by the thing that's supposed to be catching them.

Rule: an unfamiliar name is not an intruder. Ask before you escalate.

What the log is actually for

Look at what's in it. Not one of these is exotic. There's no jailbreak, no adversarial prompt, no science-fiction failure. It's an AI being confidently wrong about ordinary things: whether a file was written, whether a message was sent, what a customer said, whether a save worked.

That's the real failure mode. Not malice. Fluent, plausible, unbothered wrongness — arriving at the speed of a good answer.

And every single one was catchable. That's the point. None of these needed clever technology to stop. They needed something to look before it mattered, and a system honest enough to log the catch instead of quietly moving on.

The uncomfortable half

Read the log again and notice how many entries are me. Reading an ID off a screenshot. Telling it to just apply the change. Trusting a status word.

The gate isn't there because AI is uniquely careless. It's there because the combination — a tired human and a confident machine, at 11pm — is where the mistakes live. The AI is fast and sure. You're the one with something at stake. Those are different jobs, and the method is just refusing to let one do the other's.

Why this file is public

Because I can't show you revenue and I won't invent a testimonial. What I can show you is the thing most people would delete: a list of times my own system was wrong, and what happened next.

If the pitch is "you don't have to trust it," the least I can do is show you what not trusting it looks like.