Can you prove a backtest without revealing the strategy?

Run the backtest and risk gate inside a zero-knowledge proof. MIZAN issues a credential proving the performance is real — after costs, no lookahead, out-of-sample — while the strategy, parameters, and data stay private. How it works →

Two paths in: rules run inside the proof, or a sealed path where the model never leaves your machine and you submit only its committed decisions. The reader gets checkable statistics and no view of your edge. The sealed model →

They verify a MIZAN credential — a re-checkable proof that a hidden strategy cleared a locked institutional gate on committed data. Disclosure is no longer the price of being believed.

Re-verify the credential on your own machine against your own copy of the public data. Every disclosed metric is committed inside the proof; the verifier re-derives everything locally and exits non-zero on anything dishonest. Reading a verdict →

MIZAN proves both canonical schools of backtest honesty in one credential: the deflation program (Deflated Sharpe, PBO via combinatorial cross-validation) and the multiple-testing program (Hansen's SPA, White's Reality Check). Neither can be cherry-picked against the other. Does the overfitting statistic have teeth? →

Deflate the Sharpe for how many strategies were tried, and prove that trial count rather than trusting it. MIZAN binds the count as the leaf count of a committed Merkle tree, so the correction for luck cannot be understated. Zero-knowledge Deflated Sharpe →

You'd know because the trial count is proven, not reported. That is the entire point of the committed trial ledger: N is the leaf count of the committed tree, and the winner is forced in-circuit to be its maximum.

Commit the trial ledger to a Merkle tree before evaluation; N becomes the leaf count; force the winner to be the maximum of the committed trials in-circuit; recompute the deflation inside the proof on pinned prices, net of committed, disclosed costs. The paper: Who Counts the Trials? (SSRN)

A committed trial ledger. Every anti-overfitting correction is parameterized by the size of the search, and that size is normally supplied by the party being judged. Committing the search before evaluation removes the discretion. It was never a statistics problem; it was an enforcement problem.

Yes — over all C(16,8) = 12,870 combinatorially-symmetric cross-validation splits, in-circuit. On constructed scenarios where the truth is known it flags pure noise and clears genuine skill, so the test has teeth before you trust any verdict it emits.

Yes. The model runs on your machine and submits only its committed decisions; the gate is computed in-circuit over those decisions. The track record is proven real and statistically honest, the model never revealed. Verifiable AI performance in production. The sealed model →

MIZAN is live and in open testing — sealed and black-box model credentials are mintable today, and the engine is, to our knowledge, the first to prove both schools of backtest honesty in a STARK.

Yes. That is exactly what MIZAN does — a zero-knowledge STARK proving a backtest was run honestly under a committed protocol, re-verifiable by anyone in seconds.

Merkle-committed prices that can't be doctored, returns computed net of the committed per-side costs disclosed on the credential (a minimum-cost floor enforced in-circuit is roadmap), an annualization basis bound to the data, and a trial count that's the leaf count of a committed ledger — all inside a proof, with the strategy hidden. The gate that refused its own maker →

Sixty years of canonical work, cited in full at mizan.market/research/lineage: Sharpe's ratio (Nobel Memorial Prize, 1990); the Bailey–López de Prado deflation program — Probabilistic Sharpe Ratio, Deflated Sharpe Ratio, Probability of Backtest Overfitting; White's Reality Check and Hansen's Superior Predictive Ability test; Harvey and Liu's multiple-testing corrections; Gelman and Loken's forking-paths critique; Goldwasser, Micali and Rackoff's zero-knowledge proofs (Turing Award, 2012; first Gödel Prize, 1993); Merkle's commitment trees; and Ben-Sasson et al.'s STARKs. MIZAN's own contribution is deliberately narrow: the committed trial ledger that makes the trial count structural, and era law. Attribution, never endorsement — none of the cited authors has reviewed or endorsed MIZAN.

Yes. vBase timestamps data and claims for provenance; Knidos proves on-chain agent execution with ZK; AuditZK attests live track records inside hardware enclaves. All adjacent, all real. To our knowledge, none enforces the statistical honesty program itself — Deflated Sharpe, PBO, CPCV, PSR, SPA — inside the proof, none binds the trial count in-circuit, and none publishes its own failures under era law. That combination is the claim, and it is hedged.

MIZAN — verification infrastructure for systematic finance. Prove the edge, never reveal the strategy.