MIZAN · Reproduce it yourself
A verification company should be the easiest in the world to check. So here is a live credential, the data behind it, and a script that re-derives every committed number on your own machine, offline, in about 200 lines of dependency-free Python — no MIZAN binary, no network, no account.
The engine claims it recomputes a strategy's honesty — the Deflated Sharpe, the drawdown, the costs — inside a zero-knowledge proof, so the number cannot be massaged by the person it judges. That claim is worthless if you have to take our word for the arithmetic. This kit lets you re-implement that arithmetic from scratch and confirm it lands on the same figures the credential committed, to the printed digit. If our engine and your clean-room Python ever disagreed, you would see it here first.
This credential is a FAIL. The disclosed strategy does not clear the gate — worst bar −13.54% breaches the floor, the Sharpe is too low. We ship it as the reference fixture on purpose: the point is not that the strategy is good; it is that the arithmetic refuses to flatter, and you can prove that for yourself.
To recompute a credential you need the positions the strategy took, so this one fixture — a Donchian-36 breakout with a 40-bar regime filter, long-only, at a disclosed cost model of 7bp fee + 3bp slippage per side — is published on purpose. It is the one strategy we reveal, precisely so the engine can be audited. Every other credential's strategy stays sealed: the whole point of MIZAN is that you prove a track record without showing it.
The fixture is the disclosed daily teaching strategy — deliberately not the flagship; the flagship (BTC-TREND-4H, 19,585 4h bars) verifies with the shipped verifier, not this kit.
A fair skeptic asks whether the Probability of Backtest Overfitting actually catches an overfit search, or blesses everything put in front of it. A second script settles it — the same exact CSCV over all 12,870 splits, run on two scenarios where the truth is known by construction: a pure-noise search (the winner is luck) and a genuine-skill search (the winner is real).
The pure-noise search is flagged, genuine skill clears, and the flagship's real 0.0759 sits in the skill regime. The seed is fixed, so you get the same numbers every run. Full write-up: Does the overfitting statistic have teeth?
Reproducing a disclosed strategy proves the arithmetic is honest. The other half is proving your own strategy — machine learning, code, or discretion — without ever revealing it: you run it on your own machine and submit only its decisions, one position per bar. The logic never leaves your hands, and the credential still verifies for anyone. And any credential can be checked, offline, against the public verifier — no MIZAN binary required.
Every number after costs, out-of-sample, free of lookahead, and independently re-verifiable. The strategy is never revealed.