Permanent notes on the cryptographic verification of trading track records, the mathematics we enforce, the failures we publish, and the exact boundary of what a proof can and cannot establish. Not a blog. A record — published weekly. A new note every week, a paper roughly every month, each at a permanent URL forever, and nothing ships without a receipt: an unarmored week is a skipped week.
Written and maintained by Mohammad Muavia, Founder of MIZAN. Each note carries a permanent citation and, where empirical, a worked example you can recompute yourself — the numbers in a note about verification are themselves verifiable.
One research program, published in parts: each paper one enforceable claim, each with its priority timestamped to Bitcoin before release. The capstone — a consolidated reference-implementation paper — comes last, after each part has been public, attacked, and survived.
The manager's discretion, removed: N becomes the leaf count of a committed Merkle tree, the winner is forced in-circuit to be the maximum, the deflation recomputes inside the proof. The paper behind everything on this page. → ssrn.com/abstract=7187038
Proving a hidden model's track record without revealing a parameter: committed position streams, forward-only tracks, fail-closed semantics. The disclosure question, closed. Companion note below (№ 06).
The issuer's discretion, removed: eras named by their hash, superseded never revoked, one byte is a new judge, weakness disclosed rather than rewritten — with the production scars documented, and the extension on-chain.
Ahead · Paper IV — the refusal study: one pre-registered configuration against 34 US stocks and its own maker, as data · then the consolidated reference-implementation paper. Announced when timestamped, published when ready — in that order, always.
Bailey & López de Prado gave the discipline a way to tell skill from luck: deflate a Sharpe by how many strategies you tried. But a self-reported deflation can be gamed, the person computing it chooses the trial count. We compute the Deflated Sharpe inside a STARK, on committed data, with the trial ledger cryptographically bound, so the correction cannot be gamed by the person it judges. With a worked example you can recompute yourself, and the boundary of the claim.
Read the note →Don't trust our verifier — download ~200 lines of dependency-free Python and the pinned data, run it offline on your own machine, and watch a live credential reproduce to the digit. Both halves are checkable: the arithmetic (recompute the flagship) and the overfitting test (exact CSCV over all 12,870 splits).
Open the kit →A both-sides demonstration: the Probability of Backtest Overfitting must catch a known-lucky search and clear a known-skilled one — or it's theatre. Exact CSCV over all 12,870 splits, on scenarios where the truth is known by construction. Pure noise lands at a coin-flip 0.442; genuine skill collapses to 0.001; our live credential reads 0.0759.
Read the note →The exact line between what a MIZAN credential proves and what it does not: backtest seal versus anchored forward track, a coverage map of the strategy universe — thirteen classes provable today, three on the roadmap to ~90%, one refused on principle, and why a credential is only worth what its issuer refuses to sign.
Read the note →A verification layer is only credible if it publishes its losers as readily as its winners. MIZAN mints real cryptographic proofs of strategies that failed the gate, the first GPU mint among them, and anchors them beside the passes. Why the honest failure is the most important credential we issue.
Read the note →MIZAN's technical contributions, written as claims you can check rather than adjectives you have to trust: the anti-overfitting framework enforced in zero-knowledge, a committed trial ledger that closes “who counts the trials,” both schools of backtest honesty proven in-circuit, era-pinned and self-failing credentials — and, stated with equal care, what we did not invent.
Read the note →The other half of the problem. Proving a strategy is honest is worth nothing if proving it means showing it. The sealed model lets a quant run any strategy — ML, code, discretion — anywhere, submit only its bar-by-bar decisions, and prove the performance while the logic never leaves their machine. With the honest limits of what a seal does and does not establish.
Read the note →We ported our own trend engine to the S&P 500, ran it through our own locked gate on a survivorship-free US equity data, and the gate said no — twice: first refusing our mis-declared annualization basis, then the strategy itself. The full cryptographic receipt, downloadable and timestamped to Bitcoin, and why a refusal is the strongest credential an issuer can publish about itself.
Read the note →Seventeen checks, 81 milliseconds, one word. A line-by-line read of the live PASS credential by its own issuer — the Merkle-pinned data, the locked gate, the sealed strategy, the out-of-sample decay shown rather than buried — and, at equal length, what a verdict does not tell you.
Read the note →Straight answers to what quants, allocators, and researchers actually ask — how to prove a backtest is real without revealing it, enforce the Deflated Sharpe cryptographically, verify a track record trusting no one, prove an AI model's returns without exposing it. Every answer links to a proof you can re-verify yourself.
Read the answers →