There is a moment, the first time you verify a MIZAN credential, when the terminal goes quiet and then prints a single word. This note walks you to that moment line by line, using the live PASS on the wall — including, at equal length, what the word does not tell you.
Before that word, you have run cryptography against a file you downloaded from a stranger. After it, you know something about the world that no pitch deck could have told you. Everything below uses a real, public credential — and you can repeat every step of it, alone, trusting no one.
The credential is BTC-TREND-4H. Gate: PASS. Minted 2026-07-28 on an H100 GPU, under proof era v11 — guest image id beginning 3ac3b10b. The bundle is 225,208 bytes — two hundred nineteen kilobytes, roughly, and inside it a succinct zero-knowledge proof of an entire backtest. It verifies on your machine in about 81 milliseconds.
Everything you need is public: the bundle, the credential card, and the full dataset CSV. The registry lists this credential among 23. You do not need my permission, my server, or my goodwill — you need the era verifier, the bundle, and your own copy of the CSV. The command shape is exactly that: verifier, bundle, your CSV. It exits non-zero on anything dishonest. That exit code is the whole product.
Start where the proof starts. The dataset is BTC/USDT 4-hour OHLC, full history, Binance-derived, 2017-08-17 through 2026-07-27: 19,585 bars — every bar the pair has, not a flattering window. The bars are Merkle-committed, and the root is pinned to a canonical allowlist embedded in the verifier binary itself.
Read that twice, because it is the clause most credentials elsewhere quietly lack. I cannot hand you a doctored CSV. The verifier recomputes the Merkle root from your copy of the file and checks it against the root committed inside the proof, then checks that root against the allowlist baked into the binary. If I trimmed the 2018 bear, or smoothed a wick, or started the history in 2020 when everything went up — the root changes, the check fails, exit non-zero. The data is not "trust me." The data is arithmetic.
Every figure below is committed in-proof — not printed on a card, proven:
Notice what kind of numbers these are. Not just the flattering ones. The drawdown is there. The worst bar is there. The bars-in-market ratio shows a selective trend strategy — in the market a little over a quarter of the time — not a leveraged buy-and-hold wearing a costume. And the out-of-sample Sharpe is lower than the full-window Sharpe: 0.84 against 1.41. The credential shows you that. I will come back to it, because it is the most honest line on the card.
Numbers alone are a Rorschach test. The gate is not. The policy is locked and identical for everyone: Sharpe ≥ 1.20 · CAGR ≥ 20% · MaxDD ≤ 25% · no single bar worse than −12% · at least 3 trades · at least 20 bars. I did not choose this bar after seeing the result — the verifier checks that the gate in the proof is the canonical one; a prover cannot substitute a friendlier gate and still verify. BTC-TREND-4H clears every rail.
Costs are enforced in-circuit: 7 basis points fee plus 3 basis points slippage per side — a protocol floor. A prover can commit higher costs, never lower. There is no gross-of-fees number anywhere in this system, because the circuit will not compute one.
When you run the verifier, seventeen checks print in order. Each is a distinct question with teeth: the guest image id matches the trusted program (the proof was made by the exact code you expect, not something resembling it) · the zk receipt verifies · the committed journal matches the credential, so no valid proof can be swapped onto a different card · the data root recomputed from your CSV matches · that root is pinned to the canonical allowlist · the gate policy is canonical · the dataset is asserted canonical in-proof · the credential format is current · the exposure card reconciles — bar accounting, bands, the leverage rail · the regime policy is pinned, so the prover cannot choose buckets that flatter the story · the regime panel is internally consistent · the regime bar counts re-derived from public data match, so you can count them yourself · the annualization basis is bound to the data — 2,183 periods per year, dictated by the committed window, not declared (check C1) · committed costs sit at or above the floor · the GATE VERDICT itself is zk-attested — inside the proof, not a label on the outside · a sealed strategy commitment is present · and the credential digest is tamper-evident, end to end.
Eighty-one milliseconds. Seventeen answers. That is the read.
The strategy is never revealed. Not to you, not in the bundle, not recoverable from the proof. The credential carries a sealed commitment to the strategy specification — the verifier confirms the commitment exists and that the specification itself is absent from the proof. This is the entire thesis of MIZAN in one design choice: the economics of an edge require secrecy; the economics of allocation require proof. Zero-knowledge is the only tool that gives you both without asking anyone to trust anyone.
The wall this credential hangs on is not a highlight reel. It carries honest-FAIL credentials, and they verify by the same seventeen checks. One of them is mine: my own S&P 500 strategy, Sharpe 0.41 against the 1.20 bar. On that credential, sixteen lines pass and a single line runs red — "GATE VERDICT is PASS" is false. That is the honest-FAIL semantic: the proof just as valid, the verdict just as attested, and the verdict happens to be no.
A registry that only ever says yes is an advertisement. A registry that proves its own refusals is an instrument. When you see 23 credentials and some are failures the makers chose to publish, you are looking at a selection process you can actually reason about.
Now the section I would want first if I were on your side of the table.
A PASS says nothing about future returns. It is a backtest seal — history made honest, not foresight. Whether the edge persists is a question no cryptography can answer. It says nothing about capacity or fill quality beyond the modeled costs: seven and three per side is a floor, not a market-impact study; at size, in stress, your fills are your problem.
And look at the out-of-sample line. OOS 0.84, against 1.41 full-window. That decay is real, it is committed in-proof, and you should notice it. Most performance marketing would bury it. This credential is built so it cannot be buried — the number is proven alongside the flattering one, at the same 81-millisecond cost to check. What you do with a 0.84 is allocation judgment. That it is 0.84, and not a story about 1.41, is the credential doing its job.
A verdict is a floor of honesty, not a ceiling of insight. It removes exactly one class of doubt — did this actually happen, as stated, on real data, net of costs, against a fixed bar — and leaves every other question where it belongs: with you.
Download the bundle. Download the CSV. Run the verifier. Watch seventeen lines resolve in the time it takes to blink, and read the verdict knowing no one — including me — could have made it say anything else.
Every artifact referenced is live: the bundle, the credential card, the dataset, and the registry. The verifier re-derives everything locally from your copies. Prove the edge. Never reveal the strategy.