A compliance team had been running an AI assistant to triage incoming case files for three months, and every week the reported accuracy held steady — high nineties, the same number every Friday. Nobody had touched the prompt. Nobody had touched the guardrails. And yet, when a senior reviewer finally sat down and reread twenty of the “cleared” files by hand, four of them were wrong in ways the system had never flagged — not close calls, plainly wrong.

Here is what had actually happened. The assistant’s instructions, its tools, and the team’s own growing pile of prior examples had been sharing one bounded space for three months. Every week someone added another edge case to make it “smarter.” The bucket never said stop. It just quietly started compacting what was already there to make room — squashing the fidelity of older guidance to fit the new — and the team kept reading a steady accuracy number because the system was still confidently producing an answer for every file. It just wasn’t the same system that started the quarter.

This is context rot, and it is the reason the three verification questions exist: what is it actually reading, does it say the same thing twice, and what did it not look at. Run those three questions against last month’s “cleared” pile and you will usually find your own version of this — not because anyone was careless, because access to information was mistaken for attention to it.

A federal leadership session, thirty-some people, one afternoon. The story above is a composite of two different rooms that hit the same wall inside the same month — details changed, the wall didn’t.

The Three Verification Questions moved from Rev A to Rev B this issue — a second room, survived intact, no rewording needed.

This week: pull ten of your team’s most recently “cleared” AI-assisted items. Ask the three questions out loud, in the room. Count how many make it through unchanged.