You can’t grade your own homework. We’ve had that on the site as a slogan for a while.
It turns out there’s a proof.
What the paper says
The Verification Tax: Fundamental Limits of AI Auditing in the Rare-Error Regime, by Jason Z Wang, published 2026 on arXiv as 2604.12951.
Its Theorem 7 is called Self-Verification Impossibility. Stated as the paper states it: let z(x) = g(x,θ) be any auxiliary signal derived from the model. Then any estimator that does not use true labels satisfies
sup over Δ of E[ |V − ECE(Δ)| ] ≥ min(L, 1) / 2The proof is short enough to follow without the notation. Take two worlds: one where the model is perfectly calibrated, and one where it is miscalibrated by the maximum amount. In both worlds, every label-free observation is identical. Same outputs, same confidence scores, same internal signals. A label-free estimator therefore returns the same number for both, so it is wrong by at least half the gap in at least one of them.
Two corollaries do the work that matters for AI evaluation.
Corollary 8, LLM-as-Judge Impossibility. If a judge model evaluates a subject model using only inputs and outputs, with no ground-truth labels, its error is at least 1/2 regardless of the judge’s capability or the number of evaluations. The paper’s phrasing: “calibration without ground-truth labels is information-theoretically bounded by a constant.”
Corollary 9, Hallucination Detection Wall. Factual accuracy is a special case of calibration, so the same bound applies. Self-consistency, meaning sampling several answers and checking whether they agree, incurs worst-case error of at least 1/2. And the line worth sitting with: “Self-reported hallucination rates have zero information-theoretic content.”
There is a related result. Theorem 11 is called White-Box Irrelevance. Reading the model’s weights does not escape the bound. Model access only helps if it narrows the space to a low-dimensional parametric family. Opening the hood is not the same as having ground truth.
What it does not say
This is the part most people will skip, and it is the part that decides whether the result is being used honestly.
It is a worst-case bound, not a claim that every evaluation is wrong. The paper says so directly. In its own empirical work, the pairs that showed a significant gap “involve extreme miscalibration (ECE>0.6).” A judge model that agrees with human raters most of the time is not being described as useless. It is being described as unbounded, which is a different thing.
It does not say labels are always available. Most of the interesting questions about a deployed AI system do not have a clean ground-truth label sitting next to them. That is the actual hard problem, and the theorem does not solve it.
It is one paper. One author, published on arXiv. There is a counterpoint, Algorithmically Establishing Trust in Evaluators, which claims a rigorous method for trusting an LLM-as-judge without labelled references using a protocol borrowed from interactive proof systems. We have not reproduced it and we are not going to pretend it settles the question. The honest position is that a strong impossibility result exists, and a claimed qualification exists, and the two have not been reconciled in public that we can find.
Unverified, and marked as such: we have read the theorem statements and the proof sketch at the source. We have not re-derived the proof, and we have not checked the empirical section against its data.
Why this lands on AI evaluation specifically
The bound is general. It applies to any estimator without ground truth. But the AI evaluation industry has, over about three years, rebuilt itself almost entirely inside it.
Ask a model to grade another model’s answer. Ask a model whether it is hallucinating. Sample five responses and call the agreement a confidence score. Have a “verifier” model check a coding agent’s work. Run a benchmark where the model’s own output is the only signal about whether the output was good.
Every one of those is label-free, and every one sits under Corollary 8 or 9.
This is not a criticism of the people building them. Labels are expensive, and at the scale frontier models are evaluated, they are often impossible. The methods exist because the alternative was nothing. The problem is that the field has largely stopped saying which side of the line it is on.
There is an older result worth putting next to this one. Rice’s theorem, from 1953: no algorithm can decide any non-trivial semantic property of a program. That is a floor on what static analysis can do. It is also, read plainly, the reason you test software by running it rather than by reasoning about it.
The 2026 bound is the same shape one level up. You cannot establish correctness from inside the system being examined. Not because the tools are young, but because the information is not there.
What removes the floor
Ground truth. That is the whole answer, and it is why the bound is interesting rather than depressing.
The moment the thing being checked comes from outside the system, the invariance in the proof breaks. The two worlds stop looking the same, because one of them produces a result that reality disagrees with.
That is what we do. A system claims it fixed a bug; we run the test and see whether it passes. A model claims a benchmark gain from a fine-tune; we check whether the gain survives the counterfactual arms. A system claims an effect; we measure the effect.
The receipt we sign says what actually happened, and it says what we could not establish. It does not say the system is good, and it never says a claim is true. Our verdicts are falsified, not_falsified, certified, indeterminate. A signature is not evidence of correctness. It is evidence that a specific claim was tested in a specific way and produced a specific result.
What we are not claiming
We are not claiming our method escapes every limit. It does not.
When we cannot get ground truth, we are under the same bound as everyone else, and the verdict we issue is indeterminate rather than a guess. When the specification is wrong, a correct test of it proves nothing useful. When we run a claim on a system we cannot observe directly, we are relying on the customer’s account of their own environment for the parts we cannot see.
And the bound is about worst cases. Anyone can produce an evaluation that is right most of the time. The question the theorem forces is whether you can say how right, and that question has no answer without something outside the system to check against.
That is not a limitation we are working around. It is the reason the company exists.
The theorem is not ours. The paper is arXiv:2604.12951, and if you are evaluating AI systems without labels, it is worth an hour of your time.