VERIFIABLE AI · MIZAN-NET-2 · NVDA (4-8-1 MLP, ReLU)
FAIL · +895% but 38% drawdown, Sharpe 1.02 < 1.20 — publishedMIZAN · Verifiable AI — model inference proven in-circuitengine era · v11 · both-schools validation-complete
A cryptographic credential on the MIZAN registry — every number below is committed
inside a zero-knowledge proof and re-derived by an independent verifier. The strategy itself is
never revealed.
This credential reports its true verdict. MIZAN publishes failures with the same cryptographic care as passes — a machine that can only say yes proves nothing. The verdict above is exactly what the proof attests; nothing was retouched.
STARK sealverified — no trusted setup, re-checkable by anyone
Committed journalbound to this credential — one changed byte and every verifier rejects it
Engine erarecognized — v11 · image 3ac3b10b7930 (append-only era registry)
In-circuit inferencethe model's forward pass runs INSIDE the proof — every position is recomputed from the committed weights, so no decision series can be hand-picked
Weights commitmentsealed under the zkml weights domain — the model is proven, never disclosed
Honest boundarycompact-model path (≤100k params); large-net inference is an open research frontier and is NOT claimed
Gate policypinned — the same in-circuit gate every credential faces; returns net of the committed, disclosed per-side costs (in-circuit cost-floor enforcement is roadmap)
Lookaheadstructurally impossible — features are causal (close ≤ t), signal t → return t+1
Independent re-verificationpassed on the standalone era verifier
The strategy itselfnever revealed — sealed by construction
Committed metrics · re-derived by the verifier
Model4-8-1 MLP, ReLU (compact)
Parameters49 (compact ≤100k)
Features (in-circuit)[1, 5, 21, 63]-bar trailing returns
Decision ruledead-band 35 · long-only · lev 1×
Weightscommitted + sealed — never revealed; every position recomputed from them
Sharpe · full / OOS1.02 / 0.90
CAGR35.5% (net +895% · 7.6y)
Max drawdown · worst bar−38.1% · -12.6%
Calmar · exposure0.93 · 46.7%
Trades229 (888 of 1900 bars)
Independence (1−R²)50% · mkt 49% mom 30% rev 27%
Costs in-circuit7+5 bps
What this proves · what it does not
Provesthe committed computation ran on the pinned data under the locked gate — net of the committed, disclosed costs
Does not provefuture returns · deployable capacity · real-world fills beyond the committed cost model · transfer to other assets or windows · robustness to search multiplicity — a separate credential class, and one our own flagship fails
Trial count Nnot bound on this credential — the ledger binding is carried by deflated-Sharpe and validation credentials
Holder recordnot opted in — this holder has not elected to publish their complete record
The verdict on this page is not an assertion by MIZAN — it is recomputed on
your machine from the proof bundle. Download the bundle, grab the pinned era verifier and
the canonical dataset from the Verify page, and run:
Minted by engine era v11 — append-only era registry: a credential verifies against its own era's engine id forever; superseded ≠ revoked. How the Standard works →
ميزان MIZAN — the verification standard for systematic finance · Mizan Verification, Inc.
Every claim on this page is re-derivable from the proof bundle by anyone, trusting no one — including us. Prove the edge. Never reveal the strategy.