Methodology · one page
Five adversarial gates. One binary verdict.
A verdict issued by Nyalai is either CALIBRATED or NOT-CALIBRATED. Submission alone does not guarantee that a verdict can be issued. A verdict found to rest on errors may be withdrawn through a public correction notice; withdrawal is not a third validation outcome. Each gate uses the statistical framework appropriate to its specific question. A strategy that clears all five receives a CALIBRATED verdict. A strategy that fails one is refused.
The problem
Where strategies operate: the fourth quadrant.
In the taxonomy Nassim Taleb published in 2008, decisions divide along two axes: whether the payoff structure is simple or complex, and whether the underlying randomness lives in thin-tailed or fat-tailed distributions. Quantitative strategies operate in the complex-and-fat-tailed region: nonlinear payoffs, unbounded loss possibilities, regime shifts. Our doctrine argues that this is where statistical inference is hardest and model failures most costly. Nyalai's position is that systems operating there warrant an additional adversarial layer, external to the party that built them. That is Nyalai's position, not a regulatory requirement. Supervisory guidance on model risk management, in the United States SR 26-2 of April 2026, which superseded SR 11-7, is written for the models banking organizations run.
The five gates
Each gate uses the statistical framework appropriate to its question.
Modular validation methodology, as Joseph Simonian framed it in the Journal of Financial Data Science in 2020. Each gate below cites the peer-reviewed methodology it operationalizes. Every threshold has a stated rationale. Every deviation from the standard method is flagged in the verdict-form.
Detects strategies whose observed edge does not survive correction for multiple testing, non-normality of returns, and length of the observation window.
Detects strategies whose profitability is concentrated in a single market state and would deteriorate under an unencountered regime.
Detects strategies whose significance was inflated by directed exploration during development, the failure mode that machine-learning workflows produce by default.
Detects strategies whose stated confidence estimates diverge from realized outcomes. A signal that is directionally correct but miscalibrated breaks position sizing.
Detects strategies whose expected edge is positive but whose variance is large enough that the strategy may not remain profitable under realistic path variations.
Deployment vocabulary
Model Trust Levels. From MTL-1 Untested to MTL-5 Live.
A shared vocabulary for how much trust to place in a probabilistic prediction system. Numbered ascending intensity. Each level defined by empirical passage of specific gates. Strict nesting. Binary per level. Publicly contestable. Time-bounded.
Every MTL attribution is publicly contestable through the audit trail supplied with the verdict-form. A third party may challenge the attribution, and may request access to the code and to the frozen input to reproduce the computation.
What allocators receive
A verdict-form structured as a medical report.
Findings first, methodology available on request, references cited for audit. Two layers: the upper layer states the verdict in institutional vocabulary; the lower layer supplies the technical evidence. The upper layer never overstates what the lower supports.
Verdict scope statement
Regime scope, sizing envelope, portfolio-construction assumptions. Deployment outside the envelope invalidates the verdict. A verdict on a signal is not a verdict on the portfolio built from that signal.
Triangulation panel
Risk category times business context times regulatory frame. Adapted from the three-lens framework Senthil Kumar established while Chief Risk Officer at BNY Mellon. A verdict against one triangulation cell does not extend to another.
Communication layer
Upper allocator vocabulary (shortfall, drawdown, liquidity, embarrassment, per Matt Bank's Four Horsemen framework at GEM). Lower technical evidence (DSR, BSS, ECE, bootstrap CIs). Both required.
Citation guidance
The verdict-form is designed to be citable by model owners, independent validators, and internal audit across all three lines of institutional defense. Framework, not model, following Roland Stamm at Acadia.
Reproducibility appendix
Doctrine version applied, ʼCɩcɛ validator commit hash, statistical package versions, random seeds, bootstrap parameters, HMM state counts, data provenance. Access to the code and to the frozen input is granted on request.
Style drift trigger
Any material change in features, instrument universe, risk parameters, or claimed problem domain invalidates the previous verdict. Re-validation required, not optional.
Input integrity
Sequence commitment.
Nyalai never requests permission to publish a case, or to use it in any downstream decision, before a verdict is rendered. Consent for publication or downstream reference is collected after the verdict is rendered, on a document the submitter reviews. Refusal of downstream use does not alter the verdict itself.
1. Submission is frozen first. Timestamp and hash are recorded before any conversation about downstream use begins.
2. Gates run without stake exposure. The strategy author is never told, during evaluation, what downstream consequences might attach to the verdict.
3. Consent comes after the verdict. Publication, private client disclosure, aggregate research, or external reference are all conditioned on consent collected after the verdict is fixed.
This is not a mental-state promise. It is an order-of-operations commitment that Nyalai controls and any third party can audit. Per Refusal Doctrine section 12.4. A discipline of validity, not virtue.
The gate specifications are public.
The gates and their thresholds are documented on the Refusal Doctrine page, an editorial abridgment of the text sealed as v0.3.3.1, published under CC0 1.0. The sealed text is held by Nyalai and available on request pending verification of its citation register. Anyone can reimplement the method, contest it, or diverge from it.