To our knowledge, the first verification standard where every claim is a proof. Backtest honesty is not one discipline but two — an in-sample program and a multiple-testing program — and a credential earns Level II only by satisfying both inside a single zero-knowledge proof, to our knowledge the first engine to do so. The strategy is never revealed; the verdict is recomputed in-circuit, so no fabricated curve can pass.

Coverage is the share of systematic strategies — weighted by prevalence among practicing quants — whose performance claims can be expressed and cryptographically proven by MIZAN. A verification company does not get to quote unverifiable numbers about itself, so every percentage we state publicly is pinned to a versioned note with a closure rule: covered means all of — logic expressible via the DSL or the sealed/black-box rail, a Merkle-rootable dataset for its assets/timeframe, an honest cost model for the class, and metrics + gate in-circuit.

Two honest numbers, because there are two surfaces. The engine rail is what the proving system covers; the hosted no-code app is what a user can mint without the SDK today. The app converges toward the engine number as UI ships. The remaining path to ~90% engine is dynamic cross-asset allocation and per-strike options — each named as in-build or roadmap, not claimed as shipped; the v3 step flipped point-in-time US equities to covered (test phase) on the licensed substrate.

v3 supersedes v2 (PIT US equities flipped to covered on the licensed substrate); v2 superseded v1 — which understated engine coverage by scoring the app UI's limits. We published the corrections with attribution rather than hide them. Weights are declared estimates, re-fit at ≥100 real submissions and at every era stamp. Challenges to taxonomy or weights are invited and incorporated with attribution.

Before an engine era is stamped, we run an adversarial audit against every arm of it — forged witnesses, journal-field omission, parameter tampering, rounding-direction abuse, overflow, seed and knob grinding. The most recent full-surface pass (2026-07-24, pre-v11-stamp) attempted 18+ distinct attack classes; five were real findings and were fixed, the rest refuted with recorded rationale. After fixes the core test suite runs 211/211 green. The v10 public surface was never touched — every fix landed in the unstamped v11 lane. This is a research artifact, and we present it with pride.

Two adversarial audit documents back this section — the full-surface v11 pre-stamp audit and the composed-PBO audit — each recording the fix, the test that pins it, and the refuted-attack rationale. We do not claim an audit proves the absence of all bugs; it proves we tried, hard and on the record, to break our own engine.

The restraint is the credibility. A MIZAN credential scores what happened on committed data — it does not claim the future, and it names its own residuals precisely.

A paper formalizing the cryptographic implementation of both validation frameworks — Deflated Sharpe / PBO / CPCV together with Harvey–Liu multiple-testing and Hansen's SPA, all computed in zero knowledge on committed inputs — is in preparation. We intend to submit it to a peer-reviewed venue. There is no title-as-published, no journal acceptance, no preprint id, and no co-authors to announce yet; when there are, they will appear here, traceable.

Reviewers and collaborators are invited. If you work on backtest overfitting, multiple-testing correction, or zero-knowledge proof systems, we want your critique before submission, not after. Reach us through the contact on the marketplace and homepage.

Every credential on the Registry re-verifies on your hardware, trusting no one — including us. Every methodology document is versioned and open to technical challenge; substantiated challenges are incorporated with attribution, as the coverage v1→v2 correction was. If you can find a way to pass the gate without an honest edge, or a way to make the engine claim something it should refuse, that is a contribution, and we will credit it. A verification company that cannot be checked is just another marketing claim.

Minted on your own hardware? Your strategy never left your machine — upload just the credential bundle (plus its parts/track sidecars if it has them). The studio re-runs the independent era verifier on it; only honestly-minted, gate-passing credentials join the registry. Verification is free forever — listing is the paid service (free during launch).