Every track record ever shown to an allocator was self-reported — because checking it meant seeing the strategy, and a seen strategy stops being an edge. MIZAN closes the deadlock: the record becomes a 225 KB cryptographic credential — returns after costs, out-of-sample discipline, and the trial count itself — while the strategy never leaves its maker's machine. Anyone verifies it offline in 81 milliseconds, trusting no one. Including us.
The standard is published and frozen: VTR-1 on SSRN — free for anyone to implement, including competitors.
This is the product — a track record as a document. See a full one → · Hold one yourself →
Don't take our word
This is the append-only registry — every proof the gate has minted, passes and refusals side by side. Click a row and watch it re-derive in ~81 milliseconds, offline, trusting no one. The gate refuses most of what it sees. It refused our own flagship — it's on the wall, in red.
strategy sealed → ✓ verifiedA verifier you should trust is one you can watch refuse — including refusing the most famous returns in market history, and our own.
The greatest holding of the era — refused in-circuit. Sharpe 0.89 against a 1.20 bar, with an 82% drawdown along the way. Return is not edge.
verify the proof yourselfThe world's benchmark, examined as if it were a strategy: the proof shows it is 99% market beta — nothing left once the market is subtracted. The index cannot pass a gate built for edge.
verify the proof yourselfOur best model, refused on risk — weights never revealed to anyone, including the gate. We publish our failures on the same wall as everything else. That is what a referee looks like.
see it on the wall54 of 77 credentials are refusals. The gate earns the right to say yes
by proving it can say no.
Every credential and every refusal is timestamped to Bitcoin the moment it is minted — appended to a wall that only grows, in one direction. To fabricate a verdict dated last January, you would need a January timestamp you do not have. A competitor can copy the open standard tomorrow and begin with an empty wall, three years behind — and the gap only widens.
Code clones in a year. A dated, anchored record of honest verdicts cannot be manufactured at any price. That is the moat.
The full candidate set is Merkle-committed before evaluation. The trial count is the leaf count — it can never be understated afterward. Data is committed too.
The entire evaluation runs inside a RISC Zero STARK: returns on committed data, real costs, out-of-sample split, drawdown, and the deflated Sharpe recomputed against the true trial count — in-circuit.
PASS or REFUSED — either way it is published. A 225 KB credential carries the whole examination. The strategy never leaves the maker's machine.
Offline, in 81 milliseconds, with an open-source verifier. Flip one bit anywhere in the bundle and the STARK catches it — a tampered bundle is published next to the real one so you can watch it happen.
Below: a window of the credential's 225,208 bytes. Click any byte to flip one bit — then watch the verifier below catch it. This is a rendering of a real published experiment: a tampered bundle, one bit flipped, lives next to the genuine one on /verify so you can run this for real in your own terminal.
When the record can be trusted without seeing the strategy, the whole game re-rates — and it re-rates in favour of whoever is actually right.
Eleven engine eras of production Rust on the RISC Zero zkVM. The statistics are not computed and then claimed — they are recomputed inside the proof, on committed data, by a gate the prover cannot tune. This is what a standard has to run on to deserve the word.
An AI's track record rests entirely on the word of whoever runs it — and AI made fabricated records free: one 2026 study generated 898 trading strategies for $40; zero survived an honest backtest. So we extended the standard to the machines. A compact model's inference runs inside the proof itself — every decision recomputed in-circuit from committed weights. A model of any size proves its behaviour through the sealed path: committed positions, locked gate, forward-only chain. What it did becomes checkable. What it is stays sealed.
And the first AI we examined was our own. It returned +895% on NVIDIA — the gate refused it anyway, and the verdict is public. A standard that judges machines has to start by judging its maker's.
There is a use case for every edge-maker and every liquid investor in the world — the whole systematic market, meeting at one neutral referee that holds nothing, takes no side, and belongs to none. That is exactly why every side can stand on it.
Mint a proof and a portable, verified track record is born — the strategy never leaves your machine, and it’s yours to carry anywhere.
Require the credential as the first page of diligence — see only what's proven, before the meeting.
THIS IS THE MARKET, NOT A CUSTOMER LIST · CAPITAL NEVER FLOWS THROUGH MIZAN — IT FLOWS ON PARTNER VENUES' OWN RAILS · THE REFEREE ONLY EVER MOVES PROOF
The community is a guild of the proven — membership by proof, not sign-up.
Nobody emails a PDF of returns. No allocator asks a manager to be believed. Capital moves the moment a proof clears — and every proof, pass or refused, is already on a wall that only grows. The most important number in finance stopped being a story.
A world like that needs exactly one referee. The seat is empty — and dated.
GIPS made the numbers comparable. It still takes the edge on faith.
VTR-1 makes it provable.
Systematic finance's seat stayed empty for 46 years — because its standard needed cryptography that didn't exist yet.
VTR-1 is that standard: open, frozen, free to implement — including by our competitors. Verification is free forever. The registry is append-only, so who was early is recorded permanently. The business is the referee's seat, not the toll on truth: issuance, named institutional gates, and the marketplace layer. No token, ever — a referee cannot hold a tradeable asset and stay a referee.