The other half of the problem. Proving a strategy is honest is worth nothing if proving it means showing it. No MIZAN credential ever reveals the strategy to the person reading it — that is the whole engine's promise. What differs between the paths is only how much the prover itself sees, and the sealed model is the extreme: it sees nothing but your decisions.
Every other note in this series answers one question: is the number honest? This one answers the question that makes honesty usable at all: can you prove it without giving your edge away? For most serious quants that is not a nicety. It is the whole reason a verification layer can exist. A track record you must reveal to prove is a track record you cannot afford to prove.
To be exact about the claim, since it is the whole product: every MIZAN credential is zero-knowledge to its reader. An allocator verifying your credential learns the verdict and the metrics and nothing about how you trade — whether you submitted a strategy in MIZAN's constrained language or only the decisions it made. The two paths do not differ in what the world sees; they differ in what the prover sees. In the language path the prover evaluates your logic to derive the positions; in the sealed path — the rest of this note — it never sees the logic at all, only the position each bar. Same public guarantee; the strongest private one. This note is about that maximum.
The naive way to verify a strategy is to hand it over and let someone re-run it. That is exactly what no quant will ever do — the strategy is the asset, and a revealed edge is a dead edge. So the industry settled for the opposite: take the manager's word, look at a screenshot, trust a track record that could be fabricated. Verification and secrecy were treated as opposites. You could have one or the other.
Zero-knowledge proof is the tool that makes them the same thing. You can prove a computation ran correctly on committed inputs without revealing the inputs. Applied here: prove the strategy cleared the gate without revealing the strategy.
A track record you must reveal to prove is a track record you cannot afford to prove. The sealed model removes the trade.
The mechanism is deliberately simple, because simplicity is what makes it trustworthy. Your strategy — whatever it is, a neural net, a C++ stack, a Python pipeline, even discretion — runs on your own machine, anywhere you like. What you submit is not the logic. It is the decisions the logic made: one position per bar.
The commitment is what gives the seal teeth. Your model is bound to a cryptographic fingerprint before the verdict is known. Change a single decision after the fact and the sealed track head changes, and the credential fails every verifier on earth. You cannot seal one strategy, show a good result, then quietly swap in a different one. The thing that was judged is provably the thing you committed.
A quant with a genuine edge in a proprietary model can now prove its performance — after costs, out-of-sample, deflated for overfitting — to an allocator who trusts none of them, while the model itself never leaves the quant's control. As far as we are aware, that specific combination did not exist before. It is the difference between "trust my numbers" and "check my proof."
Because the prover only ever sees decisions, it is indifferent to how they were produced. A transparent rule set, a deep learning model, an ensemble no human fully understands, a discretionary trader's calls typed in by hand — all reduce to the same object: a position stream. This is why the sealed path covers the strategies that matter most and are hardest to verify by any other means: the ones whose edge is the code. It is also, precisely scoped, verifiable AI for finance: the model's performance proven while its internals stay private. Proving the internals themselves — verifiable inference — remains a research frontier across the industry, and we claim only the half that ships. You run the model; MIZAN proves the decisions; the world checks the proof.
There is a transparent alternative for those who want it — a rules DSL that runs fully in-circuit with nothing to export. But the sealed model is the general case, and the one most quants need.
A one-shot seal proves immutability, not the absence of hindsight. It establishes that these exact decisions, compounded on committed prices at real costs, cleared the gate — and that they cannot be altered after the fact. It does not establish that the decisions were formed without looking at the future. A sealed backtest is still self-reported history.
The no-lookahead guarantee is earned forward. The circuit proves a track is forward-ordered and tamper-evident; that its positions genuinely predate the outcomes holds only when the track's commitment is anchored to an external timestamp — Bitcoin via OpenTimestamps, or an RFC-3161 authority — dated before the first bar it trades, and checked against that anchor rather than asserted by the prover. An anchored, forward-extended track is the credential an allocator should weight most; the one-shot seal is its foundation, not its finish.
Two narrower limits, stated plainly. A position stream is a lower-fidelity disclosure than the full strategy, which is the point — but it means the credential certifies the decisions, not the reasoning, and a manager who wants to disclose more can. And a pass/fail gate leaks one bit per attempt; many proofs against varying gates form an oracle, which is documented in the threat model and bounded by design. None of these is hidden, because a verification layer that oversells its own guarantees is worth nothing.
Verification and secrecy were never really opposites. They only looked that way because the tool to hold both did not exist in finance yet. The sealed model is that tool made concrete: the strategy stays with the quant, permanently, while the proof travels to anyone. Prove the edge; reveal nothing.
Verification and secrecy were never opposites. They only looked that way because the tool to hold both didn't exist yet.
The sealed-model path is live on the current engine (image id 3ac3b10b…) and in the studio today, with fractional sizing supported. Every credential referenced across MIZAN Research is real and independently re-verifiable; the strategy is never revealed.