The Back of the Card

Three months of notes said EX-03 Régimen General.

The appointment had to go through Extranjería. The form was Model EX-03. The fee was Tasa 052. Every document gathered was for that process. Every task said EX-03. The vault had notes in multiple sections, all cross-referencing each other, all saying the same thing.

Then: look at the back of the card.

“EMITIDO BAJO ART. 18.4 ACUERDO RETIRADA.”

Brexit Withdrawal Agreement. Different form. Different office. Different fee. Simpler, actually — one step at the police station instead of two steps through a separate immigration office. But completely different.

Everything documented was wrong because the vault built on one unverified assumption. The assumption: residency permit type is Régimen General. Nobody checked the physical card. I’d asked the AI. The AI asked the vault. The vault repeated the mistake with confidence. Each session added another layer of detail to a wrong foundation.

The self-reinforcing error

This is a specific failure mode. One incorrect assumption gets encoded in a note. Future sessions read that note and build on it. New notes reference the old notes. The system adds more and more correct-sounding information around a wrong core.

The vault had eight notes that all agreed with each other. They were all wrong. They were all downstream of one unverified fact.

The fix

Ninety minutes. Look up the correct process, rewrite the notes, delete the wrong tasks, add the correct ones.

The correct information was on the back of the card. Which I had. In my pocket. The entire time.

The lesson isn’t “don’t trust AI.” The AI read the vault. The vault repeated a three-month-old assumption. Neither of them looked at the card. The card said what it was, on the back, in plain text. That beat everything else in the system. It was in my pocket the whole time.

When there’s a physical document that describes its own type: that’s the source. Everything else is a copy of a copy.