Evals measure how good a model is. Nobody measures what it earned.
Investors pay hundreds of millions to know what AI does. The question that matters before capital moves is still open.
Measuring AI is already a large business. Companies that test how good a model is, or how it behaves once deployed, have raised money at valuations in the hundreds of millions of dollars. Vals AI, which benchmarks models, reported a $40 million Series A at a $400 million valuation in August 2026. Braintrust, which evaluates AI in production, reported an $80 million Series B at $800 million in February 2026. Attestable, which proves that an AI model actually ran, raised a $20 million seed in August 2026.
Investors have decided that knowing what AI does is worth paying for. Look at what all of these measure, though, and one question is missing.
Four questions, three answered
| Question | Who answers it |
|---|---|
| How good is the model on a test? | Benchmarks and evals |
| What does it do in production? | Observability and evals |
| Did this model really run? | Proof of inference |
| What did it earn with real money? | No one |
The first three are questions about the model. The fourth is a question about the money, and it is the one an allocator, a platform or another agent actually needs answered before capital moves.
Why a good eval score says little about returns
An evaluation measures a model against a fixed test. Trading results do not work that way. They depend on costs, on timing, on how much risk was taken, and above all on how many versions were tried before the one being shown. A model can score well on every benchmark and still lose money. A trading record can look excellent and still be the luck of a wide search.
The evidence is consistent. A study of 888 trading algorithms found that a strategy's historical Sharpe ratio explained less than 2.5% of the variance in its later, out-of-sample results. A 2026 study generated 898 strategies with AI for under $40, and none held up out of sample. The number a model's operator shows you is not the number that predicts anything.
Why it has been missing
Proving what a model earned runs into a problem the other three questions do not have. To check a trading record properly, you would normally need the strategy itself: the model, the weights, the rules. But the strategy is the asset. An operator who hands it over has given it away. So the checks that exist either take the record on trust, or ask for something no serious operator will provide.
Zero-knowledge proofs remove that conflict. The record can be recomputed inside a proof, from committed market data and committed decisions, net of costs, with every version the agent tried counted, while the model never leaves its owner's machine. Anyone can then check the result offline, in milliseconds, without trusting the operator or the verifier.
What the missing category looks like
This is the category MIZAN is building: proof of performance, and the check before capital moves. Its economics differ from evaluation in one important way. Evals are bought by the people who build models. Proof of performance is paid for by whoever asks to be believed, the agent, the manager or the platform, while checking is free for everyone else. That is how credit ratings work, and it is why the standards that hold the record tend to keep it.
It is also a category that says no. Of the 111 proofs on MIZAN's public registry today, 83 are refusals, and every refusal is published, including our own neural network trading NVIDIA, refused at an 895% return because its risk-adjusted performance fell short of the bar.
Where this goes
As agents start to manage and allocate money, the most important number about them will not be a benchmark score. It will be their record with real capital, checked before the next dollar moves. The evaluation market has shown how much investors will pay to know what AI does. The larger question, what AI earns, is still open.
Run a platform where people or agents decide where money goes? One founding seat per category →
Building an agent? Check one in one call → · Allocating? Verify a credential yourself →
Funding figures as reported by the companies or press (Vals AI; Braintrust via Axios Pro; Attestable); comparable categories, not comparable valuations; none of these companies has invested in or endorsed MIZAN. Wiecki, Campbell, Lent and Stauth (2016); Wuebben (2026), SSRN 7180460. Registry figures as of 20 October 2026. Not investment advice.