What a MIZAN credential guarantees, what it assumes, and what it refuses. Published so anyone can hold us to it.
This page is the plain-language standard. The formal specification — 11 sections, appendices A–E, requirement identifiers R-01…R-41 — lives at /standard-spec/. Published corrections to both live at /errata/.
A MIZAN credential is a zero-knowledge proof that a declared backtest was executed faithfully against a dataset committed up front (Merkle-bound). The verifier does not take MIZAN's word for the result, it re-checks the proof itself.
The rigor conditions are checked independently, inside the proof, on every credential:
17 conditions checked on a single credential — seventeen doors, and the verdict is the lamp above them, not one of the doors. An honest FAIL walks through sixteen green doors and stops under a red lamp: every rigor condition holds, and the sole ✗ is the verdict itself. A chained credential checks 20.
The strategy itself is never revealed. A strategy commitment is present in every credential; the logic stays with the manager, permanently.
A proof is only as honest as its boundaries. Ours are stated here, not discovered later.
The proof binds computation to a committed dataset: whatever data was committed up front is provably the data the backtest ran on, byte for byte. The authenticity of that data — that it faithfully reflects the market — anchors to a canonical dataset commitment maintained and published by MIZAN. Strengthening that binding toward independently attestable sources is on the roadmap; until then, the provenance of the underlying data is an assumption every verifier can see, not a guarantee.
The declared policy is what's checked. The proof establishes that the strategy cleared the gate it declared — after costs, out of sample, free of lookahead. Whether that bar is the appropriate bar for a given mandate is a judgment the credential discloses to the verifier; it is not something a proof can settle.
The standard is defined by what a motivated, dishonest prover can and cannot do. Verdicts below apply to every credential.
| Attack | Verdict |
|---|---|
| Swap in better data after the fact | Blocked — commitment fixed up front |
| Peek ahead of the decision bar | Blocked — ordering checked independently |
| Quietly drop costs or funding | Blocked — cost model is part of the declared policy |
| Forge the proof | Blocked — STARK soundness, no trusted setup |
| Privately re-run many variants and submit the lucky one | Addressed forward — the Deflated Sharpe credential commits every trial in a Merkle ledger; N is read from it, never declared (§04). Pre-ledger history is unrecoverable and we say so. |
| Fabricate the underlying data | Open boundary, disclosed — the provenance layer above |
Two open boundaries, both disclosed to every verifier. A system that named zero open boundaries would be lying.
A backtest can be computed perfectly — no lookahead, real costs, out of sample, and still be worthless. Marcos López de Prado named three ways a faithful backtest still lies; a second school — Harvey & Liu, Hansen — names more. MIZAN's standard is to prove the corrections of both schools, not assert them.
Run enough strategies and one clears any gate by chance. The correction is the Deflated Sharpe Ratio (Bailey & López de Prado, 2014): deflate the winner's Sharpe by the expected maximum of N trials. It only works if N is honest, and no one reports N honestly. Cryptographically closed, live: every trial is committed in a Merkle ledger, each trial's Sharpe is re-derived inside the circuit, N is the leaf count, not a number anyone types, the winner is bound in-circuit to be the maximum of the committed trials (so a decoy ledger can't collapse the correction), and the significance bar is Bailey–López de Prado's 0.95, not a coin flip. Our own flagship credential fails that bar, and the verifier says so. Verify it →
Two boundaries. The credential certifies the search you commit — pre-ledger experiments are unrecoverable, so a credential-grade N requires a pre-registered ledger, not the same request that names the winner. And since guest v10 the significance is computed at an autocorrelation-corrected effective sample size (n_eff) — serially dependent returns make the test harder, never easier.
Tune a strategy hard enough to one sample and its in-sample brilliance is a mirage. The correction is the Probability of Backtest Overfitting (PBO, via combinatorially-symmetric cross-validation): the fraction of splits where the in-sample best underperforms out of sample. Cryptographically closed, live: the CSCV engine runs inside the circuit — since guest v10 at S=16, all C(16,8)=12,870 balanced partitions, López de Prado's own published setting, with each committed trial its own STARK verified in-circuit by an aggregator proof, the pass bar is PBO ≤ 0.5, and the credential is a real STARK you can re-verify. Our own flagship sample scores PBO 0.0759, not overfit, and the verifier says so. Verify it →
Ordinary cross-validation lets information from the test period leak into training through overlapping labels and adjacent bars. The correction is purged, embargoed combinatorial cross-validation (CPCV): remove overlapping observations, embargo the boundary, and score across many recombined paths for a distribution of out-of-sample performance, not one fragile number. Cryptographically closed, live: the purged, embargoed recombination runs inside the circuit and the headline is the conservative 5th-percentile path Sharpe across the distribution, not the mean. Our flagship sample's P05 path Sharpe is +0.003 — positive across even the worst paths, and it is a real STARK you can re-verify. Verify it →
López de Prado's three corrections ask whether a winner survived its own search. A second, independent school — Harvey & Liu on multiple testing, Hansen on Superior Predictive Ability — asks a harder question: given the trials and the noise, is the Sharpe distinguishable from luck at all? Cryptographically closed, live on guest v11 (3ac3b10b…): the Probabilistic Sharpe Ratio: is the Sharpe positive at 95% confidence once skew and kurtosis are accounted for; Superior Predictive Ability: does the winner beat its benchmark after the full search is priced in, and unlike the Deflated Sharpe, SPA refuses zero-dispersion decoy trials designed to soften the correction; and minimum backtest length: is the track even long enough to support the Sharpe it claims. A credential that clears both schools is far harder to fake than one that clears either. Our own flagship's PSR reads 0.7226: below the 0.95 bar, and the verifier says so. Verify it →
Honest boundary. Minimum backtest length is currently consistency-checked verifier-side rather than fully re-derived in-circuit; the Probabilistic Sharpe and SPA are proven inside the circuit. Stated here, not discovered later.
Status, stated plainly: both schools of backtest honesty are now cryptographic and public. López de Prado's three failure modes — selection bias (Deflated Sharpe), overfitting (PBO · CSCV), leakage (purged, embargoed CPCV) — live since guest v10; the Harvey-Liu / Hansen school — Probabilistic Sharpe, Superior Predictive Ability, minimum backtest length — live on guest v11. Each is a real STARK credential you can re-verify yourself. To our knowledge, the first time the complete statistical framework — both schools — has been made verifiable, not asserted.
A credential attests that a track record cleared a declared gate on committed history. It says nothing about future performance, and never will.
The credential verifies computation over committed history, not live fills. Chained tracks shipped 2026-07-16: the same sealed strategy re-proven over a strictly extended window as data accrues, a growing, verifiable track, every link bound to its parent (the first real chain is public on the live chain). The boundary stands: chaining extends the proven window; it does not attest execution or fills. That layer remains in scoping and is labeled exactly that.
High-frequency strategies cannot be honestly verified from bar data — fill realism, queue position and adverse selection dominate the result. We refuse the category rather than bless it weakly.
PASS means the declared conditions held on the committed data. It is not investment advice, not a recommendation, and not an opinion on whether the strategy suits any mandate.
The verifier is source-available. The sample credential is real. Run it yourself. Every correction filed against the standard or its reference implementation is on the public register at /errata/ — spec errata and implementation findings, cross-linked, published whether or not they flatter us.