Confident, and wrong.

Two of them agreed

We were going back through a stack of old work, pulling out the lessons worth keeping. To move fast, we had two separate AI helpers each read the same past job and report what happened.

They came back with the same fact — a specific technical detail about how a system was set up — and both stated it flatly, no hedging. Same answer, twice, full confidence. We almost banked it.

We checked the receipt

We have a rule that a summary doesn't get to change how we work until we've checked it against the original recording. So before writing it down, we went back and read the actual record of that job.

It said the opposite. The thing both helpers reported as true was backwards — and they'd gotten it backwards the same way. Two confident answers, one shared mistake.

Confident and wrong is worse than "I don't know"

An honest "I'm not sure" is safe — it asks to be checked. A confident wrong answer is the dangerous one, because nothing about it raises a hand. It just sits there looking like a fact.

If we'd trusted "well, two of them agreed," we'd have written that backwards detail into how we work — and then repeated it on every job after this one. You don't pay to make a mistake once. You pay to make it over and over without noticing.

Agreement isn't proof

Here's the trap: both helpers read the same pile. When two summaries come out of the same source, they share the same blind spots. Them agreeing isn't a second opinion — it can just be the same error, printed twice.

So "multiple AIs said the same thing" is not evidence. It's a reason to go look at the source, not a reason to skip it.

The rule

A summary is a lead, not a fact. It doesn't change how we work until we've confirmed it against the thing that actually happened. That check takes a minute. Being wrong on your dime takes a lot longer, and you rarely find out until it's expensive.

Why this is on our website

Because the whole pitch of handing your tech to someone is that they catch this stuff. The difference between operators isn't who uses AI — everyone does now. It's who forwards the robot's confident guess, and who checks the receipt before it reaches you.

On your job, that's the difference between a clean result and a wrong "fact" that quietly compounds — because a confident mistake doesn't announce itself. It just keeps being wrong until someone bothers to look.

We don't trust the summary. We read the source. Tell us what's broken.