Show Me the Output
Stop accepting reports. Ask for the artifact — the actual numbers, the actual error, the actual file — and the confident-sounding failure disappears.
By Geordie Everitt
"I've updated the spreadsheet." "The numbers check out." "That should work now."
Each arrives sounding like good news. Each is a claim about a result, delivered in the cadence of a result. The distinction is easy to lose, because in ordinary conversation with a competent colleague there is no gap between the two. They say the file is updated; the file is updated. The report and the thing reported are welded together by the fact that a person who says otherwise looks bad on Thursday.
Nothing welds them here.
Reports are cheap in a way you are not used to
When a colleague tells you the reconciliation balanced, a great deal of invisible machinery stands behind that sentence. They have to face you at the next meeting. Their name is on it. If they are wrong twice in a quarter, something changes for them.
None of that machinery exists on the other side of a chat window. A model producing "the reconciliation balanced" is producing the most plausible next sentence given everything before it — and after an hour of cooperative work, the most plausible next sentence is very often the one where it worked. There is no reputational cost to be weighed, because there is nothing there to weigh it. The sentence is not a lie. Lying requires an interior.
This is where the habit has to change, and it is a small change. Stop accepting reports. Ask for the artifact.
The four words
Show me the output.
Not "did it work" — that invites another report. Not "are you sure" — that invites a more emphatic report, which is worse, because emphasis reads as evidence. Ask for the thing itself: the actual numbers, the actual error text, the actual paragraph, the file as it now stands.
The reframe is small and the effect is not. A model asked to produce an artifact has to produce one, and an artifact either exists or it doesn't. There is no confident version of a table that isn't there.
Two phrases to ban outright while you are at it: "should work" and "looks correct." Both describe an expectation while wearing the clothes of an observation. If you see either in a report, the check was not run.
The instruction, if you want it standing
Put this in the instructions you carry from conversation to conversation, so you stop having to remember:
Never claim something works, ran, or is fixed without showing me the actual output. Do not say "should work" or "looks correct." Run it, read the result, show me the result, then make the claim. If you cannot run it, say so plainly and tell me what you need.
That last clause carries more than it looks. Most confident-sounding failure is not fabrication; it is a gap the model papers over because papering over gaps is what plausible continuation does. Given explicit permission to report the gap, it usually will.
What you get back
The first thing you notice is fewer wasted afternoons. The second is stranger: your questions get better.
When the answer has to be an artifact, you start asking questions that can produce one. Instead of "is the pricing model sound" you ask to see the model with three real scenarios run through it. Instead of "is this draft good" you ask which paragraph carries the argument and why the others exist. Demanding evidence turns out to be a discipline on the person demanding it, which is a heavier ask than it sounds.
You do not have to become suspicious to do this. The machinery that makes a colleague's word reliable simply is not present here, and asking to see the work is not an accusation. It is the same thing a good editor does to a writer they like.