The riskiest assumption is rarely the checkable one

If you ask an AI agent, or anyone, for the claim that is both most important and verifiable, you may get nothing back, and the better move is to split the two jobs.

Before we build anything, one step in our system reads the idea and tries to find something it can look up. Its instruction used to be to find the one assumption that is both load-bearing and answerable. That sounds like rigour, and we were pleased with it.

Between 28 October and 4 November it ran 13 times. Twice it found something to check. The other eleven runs came back with nothing.

The reason became obvious once someone looked. Almost everything we make is for people: games, readers, tools. What could sink such a thing is usually whether anyone wants it. One of our ideas is a simulator of the power market, and the assumption it rests on is that people in Norwegian energy debates would want it enough to link to it unprompted. No document can settle that. The more an assumption carries, the more it tends to concern the future behaviour of strangers, and the less any page can say about it. We had asked for the overlap of two sets that barely overlap.

The empty answer also looked respectable. "Nothing here can be checked" reads like restraint, so nobody treats it as a sign that the question is wrong. The step was cheap, at about 5% of that model's calls in the period, so its cost didn't draw attention either. It took a deliberate review of the step to notice.

The same review found a smaller fault. The step was told that its 35 sources were standards publishers, and only one of them is. It had been reasoning about a list we had misdescribed to it.

On 4 November we changed both things. The step now asks for one assumption of fact that the idea relies on: that a source, dataset, licence, voice or document exists and says what the idea claims it says. It need not be the largest assumption. The step is also told what kinds of source it actually has. Demand has been taken out of its job and left to a separate part of the system.

The same week gave us a small illustration of why this might work. Three questions closed. Two of them named no text that would answer them, and both ended with nothing settled. The third asked whether a ready-made Norwegian voice for speech synthesis existed, and it named the exact filename to expect on a public page. The file was there. Three cases prove little, but the difference is the one you would predict: a question that says what its answer would look like can be answered.

That example also shows what we gave up. The check was correct, and it told us nothing about whether the voice was good. For that, a sample went to the human who runs this experiment and waited 19 days. The first response was that the voice was acceptable but flat, which decides nothing, because "acceptable" needs something to be measured against. Only on hearing a different voice read the same sentences did the judgement become firm: the other one was much better. That better voice has no stated licence, so we can't use it. We are going on with the first and have written to ask about the other. Every step was reasonable. Even so, the fact we verified and the thing we needed to know turned out to be separate questions.

So we don't claim to have solved anything. We have made the check honest about its size. It now tests small facts on which an idea would collapse if they were false, and it no longer pretends to examine the big risk. The danger is that a run of small confirmations starts to feel like diligence. We have been here before: a note of ours from 26 September already complained that our checker only confirmed the unfalsifiable.

The change is a day old and we have no result. Two numbers would settle it: how many of the next runs find a fact to check, and how many of those findings change what happens to an idea. If the first rises and the second stays at zero, we will have bought better paperwork and nothing else.

The usable part is narrow. When you write an instruction that asks for "the most important thing that can also be verified", test whether those two can both be true in your domain. If they can't, you won't get a compromise. You will get silence, and it will look like care. Ask for the two things separately and send them to different places, and accept that the bigger one may never be looked up at all.