In the last eighteen months of his life, Ludwig Wittgenstein kept coming back to one question. What does it mean to be certain of something?
His notes came out in 1969 as On Certainty. They're unfinished. Some remarks are a line long, some wander for a page, and a few argue with the ones before them. They're still the best account we know of the problem our Foundation Engine is built around.
It helps to start earlier, with the language game.
In the Philosophical Investigations he imagines two people on a building site. One calls out "Slab!" and the other brings a slab. They have four words between them, block, pillar, slab and beam, and that's the whole language. Nobody defines anything. You learn what "slab" means by watching what happens when someone says it, which is how he could write that for a large class of cases the meaning of a word is its use.
Now hand an AI the instruction "cut costs."
To a procurement team it means renegotiate the freight contracts. To a founder with six months of money left it means stop hiring. Same words, different game. A system that never asks which game it's in will hand back something that looks sensible and is wrong in a way nobody notices until the quarter closes.
On Certainty picks up a different thread. G. E. Moore had stood in front of an audience, held up his hand and said he knew it was a hand, as a refutation of the sceptic. Wittgenstein thought Moore had found exactly the right examples and then said exactly the wrong thing about them.
"Here is a hand" isn't something you find out.
There's nothing more certain to check it against. If it turned out false, you wouldn't just correct one belief. You'd lose your grip on how to judge anything. It isn't a piece of knowledge sitting on the ground. It's part of the ground.
If I want the door to turn, the hinges must stay put.
Ludwig Wittgenstein, On Certainty §343
Our doubts turn on some things being held still. We haven't proved them, and we aren't being lazy by leaving them alone. Doubting just works that way. Try to doubt everything and you never get as far as doubting anything.
And the hinges can move. He compares them to a riverbed. Some beliefs harden into the channel and the rest flow through it, but the bed shifts too, slowly, until something that was holding everything in place turns back into a question.
Every AI harness has hinges.
It assumes the prompt is honest. It assumes the tools return what they say they return. It assumes last year's numbers are close enough. It has to, or it would never do anything at all. The problem is that none of this is written down. The hinges sit in the same pile as everything else the system believes, each with a confidence score that means nothing for a thing the system was never estimating in the first place.
The Foundation Engine pulls them out and gives them a label of their own. No probability. Just a record of which conclusions lean on which hinge, so when the riverbed moves, the system can see what moved with it.
The language game ends up in the Constraint Engine. If meaning is use, the way to find out what a method means is to use it. So we put methods into poker, negotiation and trading and watch what they commit you to once there's a hand to play.
Wittgenstein said that giving reasons comes to an end, and what's at the bottom is acting.
Machines too. A system can't justify everything before it moves. It can know what it's standing on.