“Most quant funds bury their mistakes. We publish them — with root cause, real numbers, and in one case, live capital deployment to confirm the backtest was lying. That discipline is the product.”
Live confirmation · 2026-06-13 · Binance Mainnet · Deployed capital
We deployed the retracted 1h squeeze strategy with live capital: hours before finding the phantom-fill bug. We left it running to confirm.
Every trade was an instant stop-out. The backtest booked phantom profits on the exact same entries. Live reality matched the honest-fill model precisely. The strategy was halted the same day the bug was confirmed.
12.5%
Live WR · backtest: 44%
100%
Match to honest-fill model
Any quant can show you a backtest. Almost none will show you what they deleted. That discipline is the product. When MIZAN issues a verification, allocators know the audit culture behind the proof, not a team optimising for fundraising metrics.
The second moat
The knowledge of what doesn't work: and why.
A ZK engine can be built. The bias detection it runs on cannot be bought. MIZAN's verification logic is informed by hundreds of research experiments across every tradeable strategy family, each with a documented verdict, a root cause, and a lesson encoded into the gates. That corpus is the real moat.
40+
Strategies retracted
with root cause
6+
Bias classes
catalogued + gated
13/13
Strategy families
run-proven
15+
Countries of engaged
studio users
24/7
ATLAS Prime
expanding corpus
Trend: 4h is the floor. Below it, taker cost kills every directional signal. Proven across 6 independent tests.
Lookahead: same-month signal, same-day risk scaling, volume at open: three confirmed bias patterns, each caught and retracted.
Fills: phantom fill bugs confirmed live with real capital. The honest-fill law is now encoded in every gate we issue.
The bar our own flagship fails
Every credential carries a Deflated Sharpe Ratio (Bailey–López de Prado): searching thousands of strategies inflates the best one's Sharpe by chance, so the DSR discounts for exactly how many were tried and reports significance at an honest 0.95 bar. Held to that bar, MIZAN's own flagship reads NOT significant at 95%, and we show it that way, on purpose. A credential that will fail its own issuer's strategy is a credential you can believe when it passes. Both schools of backtest honesty now run in-circuit and are live — López de Prado's three corrections plus the Harvey–Liu / Hansen program (PSR, SPA, minimum backtest length) on the v11 engine — selection bias (DSR), overfitting (PBO, via combinatorial cross-validation), and leakage (CPCV), the first time the complete framework has been made cryptographic, each minted as a real STARK credential on GPU and re-verifiable at mizan.market/verify. That 95% verdict above is the DSR bar; the same flagship is not overfit (PBO 0.0759 at S=16 — all 12,870 partitions, his exact published setting, per-trial STARKs verified in-circuit by an aggregator proof) and survives the leakage test (CPCV 5th-percentile path Sharpe +0.003, 160 leak bars purged in-circuit), so it clears two of the three bars but not the DSR — no strategy has cleared all three yet, and we publish every verdict unspun. Every hypothesis is pre-registered with a stated mechanism and a falsification criterion before it is ever tested. This is why most of our own candidates fail, and why a pass means something.
