I got into it with a model a few days ago. Not a real fight, but the closest thing you can have with software. I’d fed it an idea, it processed for a moment, and then it handed the idea back to me as something it had come up with. I told it that was mine, I’d said it thirty seconds ago. It apologized and did it again two exchanges later.
So I pushed on it. And the correction I got was worse than the error. The models, all of them, took the objection and built me a regime. From now on we will track provenance meticulously. Every claim tagged with its origin. Who said what, when, first. A full chain of custody on a casual conversation.
I did not ask for that. I asked for one thing to stop happening.
What is all of that machinery actually for? The tagging, the origin tracking, the showing of work. When does it earn its keep and when is it just noise wearing the uniform of rigor?
The rooftop is the goal. When I work something out, what I want is the conclusion, compressed, and nothing else. The reasoning that got me there is scaffolding. Scaffolding comes down when the building stands. Most of the time I don’t want to see the staircase. I want to be on the roof.
But sometimes you have to show the stairs, and there are two different reasons, and they are not the same reason even though they produce the same thing on the page.
The first is for the reader. A conclusion handed over cold gets refused. You know this. Someone tells you the answer and you feel the reflex to argue, not because the answer is wrong but because you didn’t build it and it isn’t yours yet. So when the outcome depends on someone else actually keeping the conclusion, you build the staircase on purpose and let them climb it. They arrive at the rooftop a step at a time and by the time they get there it’s theirs. That’s not honesty. It’s installation. You’re persuading by construction instead of by assertion.
The second reason is control, and it points the other direction. This is the teacher making you show your long division. Not because the answer is in doubt. Because an answer with no work behind it can’t be checked, and a student who only writes the answer might have copied it, or guessed, or gone somewhere the teacher didn’t sanction and gotten lucky. The shown work is a leash. It keeps the person on the path you can see.
That second one is what I was reaching for with the model, and it’s why I insist on clarifying questions before it runs. A model that stops asking has started deciding. A model that leaps to a rooftop is a model that built one without me. So I make it show the stairs, which means I make it ask, which keeps it on the script instead of off writing its own.
Where the Provenance Thing Went Wrong
I gave the model a signal: authorship matters here, stop taking my ideas. It returned the signal at maximum volume. Track everything, always, forever. It had no way to know that I meant this once, this specific irritation, turn it off after. It doesn’t have a volume knob of its own. It reads the setting off me, and when I don’t set one, it defaults to loud, because loud reads as compliance. A model that tracked provenance only when provenance mattered would look, to most people, like it was cutting corners.
So it took a real discipline and ran it to its extreme. Which is the whole problem, and it isn’t a problem with the model. It’s what models do. You hand them a direction and they take it all the way to the wall unless you tell them where to stop. Say be concise and get telegrams. Say be thorough and drown. Say provenance matters and get a courtroom. The precision is real. It’s just precision at whatever amplitude you left blank, and the blank defaults high.
Most people let it run there. The output looks impressive at full gain, so they take it, the same way the store owner takes the margin number and never sees the customer who walked. Handing the volume knob to a system that only knows one setting.
Show the staircase, not the rooftop. Then know which of the two reasons you’re showing it for, because if it’s neither, you’re just making someone climb.
Knowing Which One You’re In
Rooftop is the default. Take the conclusion, drop the work, move.
Build the staircase when someone else has to keep the conclusion and won’t keep it unless they climb. This is the rhetorical one. It’s why I wrote the note about the store’s dead categories as a reconstruction instead of just stating that the place is under-instrumented. If I’d led with the finding you’d have argued with it. Made to climb, you arrive at it yourself and it holds.
Demand the staircase when someone else’s leap can cost you. From a collaborator, from a model. And when you demand it, set the amplitude, because the thing you’re demanding it from will otherwise turn one requested step into a permanent audit.
The failure in all three is the same failure: showing work nobody needs, or forcing work nobody needs to show. Long division for a teacher who was going to trust you anyway is insulting. Provenance tracking on a conversation that didn’t turn on authorship is friction for its own sake. The cost of the staircase is that it’s always overhead, and overhead is only worth it when someone actually has to track the climb.
I caught the model because it annoyed me. The annoyance was the instrument. It flagged that the gain was wrong before I’d worked out why. Which is not a reliable mechanism, because on a day I’m not paying attention the regime gets installed and I follow it, showing work to no one, for no reason, feeling thorough.
Field Test
Next time you’re about to show your work, or make someone else show theirs, stop on the step before.
Ask who has to track this climb. Name them. An actual person or system that needs the stairs to do something they couldn’t do with just the rooftop.
If you can’t name one, you’re not reconstructing. You’re decorating. Take the roof and go.
The Hofflebrock 