Truth Is Buildable

I keep auditing benchmarks that cannot be reproduced. DeepSWE publishes a real verifier and withholds the patches it graded, so its headline number is honest arithmetic over verdicts nobody can re-derive. Another vibe-coding bench has a language model award points against a prose checklist, so its verdicts were never deterministic. Different mechanisms, one product. A number you cannot check.

I went looking for a sentence to put on that number and found one already in my mouth: what cannot be false cannot be true.

It felt borrowed. No attributed source turned up, though a near-twin already sits in Angelo’s Wittgenstein commentary (2022), which had the tautology half before me. The full hunt is in the endnote below.

The line is a compression. Here is the decompression.

What cannot be false

Two things can be immune to falsehood, and they reach it from opposite ends.

A tautology cannot be false. It is true in every possible world, which is why it forbids nothing and informs nothing. “It will rain or it will not.” Unfalsifiable by logic, and empty for the same reason.

A number you cannot check cannot be shown false either. No procedure can come out the other way, because there is no procedure you can run. The 70 percent that cannot be re-derived is not a wrong measurement. It is not a measurement. The tautology cannot be false because it says nothing. The uncheckable number cannot be shown false because it shows nothing. A benchmark that cannot be reproduced and “it will rain or it will not” are the same species. Both survive every test by taking none.

Cannot be true

Start with the word the line leans on. Truth here does not mean correct forever. It means a claim that has stayed standing after it was given a real chance to fall. That fixes an order. The capacity to be false comes first, and the right to say true is earned only after. What cannot be false never reaches the question of truth at all.

So the immunity is the disqualification. The test for truth was always the risk of being wrong. Anything that runs no risk has bought its safety by emptying itself. A measurement that cannot fail is a tautology wearing a decimal point.

Turn that over and the equation falls out. If it cannot be false, it cannot be true, so the capacity to fail is the price of admission, and it is paid first. Spend it: a claim is true when it can be false but isn’t. Not true forever. True as standing. Truth is what could have failed and hasn’t.

Take the hardest case. Suppose I tell you I know the precise number of atoms in the universe. Call it N. Two claims hide in that sentence. One is about the world: the number is N. One is about me: I know it.

The world-claim is contingent, so it can be false, and my N almost certainly misses. But nothing can show it false. No one can count. It might, by luck, be exactly right, and it still would not be knowledge. Correct and unbuildable at once.

The knowledge-claim, I know N, you can test, and not by counting atoms. You demand the build. Show the chain, the derivation, a source that reaches a root I can check. There is none. So I know N is false, refuted at the level of provenance, while the world-claim under it stays out of reach. You could not check the number. You could check that I never built a warrant for it.

That is the machine in one example. A claim about the world can sit past every test. The claim to know it cannot, because knowing is a build, and a build either shows its chain or it does not. The benchmark has the same shape. Maybe the score is right. “We measured it” is false the moment the evidence is withheld, because the measurement was never built where anyone could climb to it.

Truth at the edge

Knowledge is a graph. A claim is a node; the citations and inferences that wire it to what it rests on and what it implies are the edges. Truth does not live in the nodes. It lives in the edges. A tautology is a single node: irrefutable, detached, wired to nothing. The irrefutability and the uselessness are one property. It has no edges.

The edges are what you grip. Provenance is the trace they leave, the handle that lets you grip a claim, push on it, and find it wanting. No handle, no grip, and no truth to be had, only assertion.

Which produces an ordering nobody expects. “Because the Bible told me so” is not being praised as evidence. It is being praised, narrowly, for naming its source. It cites its provenance. You can open the book, locate the verse, dispute the reading, weigh the authority that hands it down. It puts its source on the table and invites the argument. In that narrow sense it is not more falsifiable as a claim about the world. It is more accountable as a claim about why someone believes it. It tells you where to push.

And the citation does something precise. It makes the belief inherit the fate of its source. If the cited verse does not say it, or the authority behind it does not hold, the belief may fall too. That conditional is the falsification channel. Naming a provenance is the act of saying: here is the thing that, if it fails, takes me with it. The claim hands you a target and a way to bring it down.

It does this by moving the argument up a chain. The dispute stops being whether to believe and becomes whether the source holds, a question you can put to the source, which can be made to cite its own. Each citation is a rung. Falsifiability is the chain being climbable, link by link. Truth is not the rung at the top. It is the fact that you can keep climbing.

The uncheckable number hands you nothing to push on. Its verdict is a promise, not a trace. It cites a procedure it will not show you, so no source’s falsity can ever flow into it. It depends on nothing you can attack, which reads as strength and is the entire defect. The chain ends at the first rung, and the rung is sawn off. The argument cannot move.

On provenance, the decimal point ranks below the scripture citation. The artifact dressed as measurement is the purer act of faith, and the worse one, because faith that names its source can at least be argued with. A number that hides its source has taken “because I said so” and deleted the “I” you could have answered.

The scripture goes one better than citing outward. At its base it cites itself, and it names that base honestly: faith. That is a complete stack trace. You can follow every frame to the root, see that the root is an axiom and not a measurement, and decide for yourself whether to accept it. Nothing is hidden, least of all the bottom. The number’s trace breaks partway down. has_model_patch: true is a frame pointing at a deeper frame that is not there, a dangling pointer where the evidence should sit. Both systems rest on something you finally take on trust. Only one of them signs its name to where that is.

Reverse the instinct, then. We file “the Bible told me so” at the bottom of the epistemic ladder and “70.045 percent” at the top. They sit the other way around. The one that cites can be checked. The one that computes in private cannot.

Truth is buildable

If truth needs the capacity to fail, then truth is not received, it is built. A true claim is a structure you assemble from sources, and like any build it has dependencies, a log, and a state that is green or red. Provenance is the dependency graph. Citation is an edge to what the claim rests on. Attestation is the signed build log, the line that says I built this, here is the receipt. Falsifiability is the build being able to go red. The test is reality pushing back. Truth is the build currently passing, and reproducibility is whether anyone else can rebuild it from source.

The guardrail is the one that keeps this from being relativism. Buildable does not mean manufacturable to spec, with the building making it true. The build includes a test, and the test can fail. Truth is buildable and breakable, and the breakability is the whole point. A claim whose build can never fail is not true. It is a hardcoded return value. The number that grades itself and will not show its work is return 0.70, a mocked test reporting green.

This post says truth is built. That denies the real nothing, it only moves the real to where it always was, ahead of the build, the gap past the edge of the known. Crossing that gap is inquiry, and raising the hypothesis so the test has something to try is one of its three modes of reason, abduction.

Companion: belief is the edge of knowing comes at the same edge from cognition. There is no boolean truth above belief, only confidence past a threshold. Here that confidence is something you build, and can break.

Provenance endnote

The search behind the verdict, kept so you can rerun it and refute me.

A second model searched independently. It confirmed the verdict and surfaced Angelo, which my search had missed, one blind spot caught by another.