Methodology
The construction is protected. The proof is public.
Deterministic where it provably publishes.
Every score, every verdict, every projection from the Sovereign Risk Infrastructure is reproducible, auditable, and anchored to validation evidence – not to a black box. The mathematical method behind the ARIN22 kernel is published at the level of its mathematical class; the construction itself is a protected trade secret, disclosed only under agreement.
Proof discipline
Evidence before formula.
The publishable claim is always anchored to evidence, never to a secret construction. Public pages show the envelope; NDA packs show run evidence, hashes and replay bundles; signed engagements handle protected derivation.
Verified truth.
Every output traces to source data, snapshot, kernel version, routing class and replay hash. The result is inspectable by a model-risk reviewer.
Lane-qualified metrics.
Latency, parity and high-dimensional accuracy are always tied to a lane. D = 4 latency is not sold as D = 200 latency.
No silent degradation.
Hard regimes route to EVT / POT overlay or MC challenger lanes. The route is recorded before publication and carried into the audit trail.
Stochastic to deterministic
From path enumeration to deterministic tail metrics.
Classical Monte Carlo enumerates the future by brute force. ARIN22 reduces dependence on repeated enumeration inside bounded lanes and reproduces Monte Carlo tail metrics within published validation bands.
CVaR 99.9 within 0.035% of a 10M-path MC reference on the canonical lane; D = 200 high-dimensional lane max 0.90%. Deterministic where it provably publishes. EVT / MC challenger on hard regimes. No silent degradation.
The kernel method
What the kernel computes – and what stays sealed.
The public method describes the admissible class and the evidence envelope. It does not disclose protected construction, calibration thresholds or implementation details.
Named mathematical basis. Protected deterministic risk-state construction with EVT / POT tail overlay. The class is public. The construction is not.
Published authorship. The deterministic risk-compute method is published under the founder’s name, distributed across SSRN eJournals:
Anti-hallucination
Math governs the agents.
Risk numbers are computed by the deterministic kernel, never by an LLM. Agents explain, challenge and route. The Governor decides what can leave the system.
Deterministic constraint.
VaR, CVaR, stress and cascade numbers are reproducible from snapshot plus kernel version. Narrative cannot override the computed state.
Release-managed learning.
Realised-outcome recalibration enters through versioned calibration releases with replay tests and approval gates, never through silent runtime drift.
Single exit.
Before publication, a math-to-narrative audit blocks contradictions. The Governor is the single exit, and every verdict carries a cryptographic trail.
Latency to autonomy
Autonomy in regulated finance is a latency problem.
A council can deliberate and govern, but if risk revaluation takes minutes, no committee will trust the loop to act. ARIN22 validated lanes reduce core revaluation into sub-millisecond bands and move toward the full-stack 400 ms design target (Lane 3) – the measured value is published at contract signing, not before.
Execution remains client-side. SAA is not a broker, venue or execution engine. Precise benchmark configuration, speedup multipliers and compute-economics models are released to qualified partners under NDA.
Disclosure boundary
Public envelope. NDA evidence. Protected construction.
The page is intentionally explicit about what is public and what is not. Serious diligence gets evidence. Public readers get the claim boundary.
Public surface.
Latency envelope, accuracy bands, validation counts, class routing, regulatory mapping, integration surface and reproducibility contract.
Trade secret.
Protected construction, construction dossier, fallback logic, parameter calibration and kernel implementation.
TIER 1 – public page · TIER 2 – NDA pack · TIER 3 – signed engagement · No third-party validation attestation is claimed until issued.
External evidence
External evidence
Published replay lanes, public-source arithmetic and withheld boundaries in one diligence ledger.
External evidence means we replay published figures from public sources through our kernel and publish the receipts. No regulator, index provider, or data vendor has reviewed or endorsed SAA. Our own loss models are withheld from these exercises by design.
| Lane | Public source | Result | Status |
|---|---|---|---|
| Changepoint detection (TCPD) | Turing Change Point Dataset, 15 methods | 4/15 mean-score (CI95 1-7) · earlier successful-rows frame 6/15 lies inside the CI – frames reconciled | PUBLISHED · REVIEW_ONLY |
| DFAST 2025 capital arithmetic | Federal Reserve, 22 banks | max error 0.000000 | PUBLISHED · REVIEW_ONLY |
| EBA 2025 capital arithmetic | EBA EU-wide, 64 banks | max error 0.000000 | PUBLISHED · REVIEW_ONLY |
| DebtRank replication | BIS LBS, EBA bank-sovereign, SEC 13F (4-quarter stability), FFIEC NIC (topology only) | six networks, receipts per sheet | PUBLISHED · REVIEW_ONLY |
| Crisis VaR/ES backtest ladder | 1,972 rows across public crises | 201 failures vs 49.3 expected – formal ES research rejects production validation | SELF-REJECTED · predates Aug 2026 kernel revision · re-run scheduled post-productionization |
| NYU V-Lab SRISK comparison | V-Lab method pages | not run | WITHHELD · license boundary |
| Own loss models (Level 3) | – | not run | WITHHELD · by decision |
Internal acceptance evidenceInternal acceptance evidence is separated from external replays by design. Estimated-config acceptance battery vs an internally generated 100M-path reference: 160/160 cells within the 5% envelope across 10 market regimes · median |error| 0.0156% · parameters hash-frozen before every comparison · every intermediate failure published · REVIEW_ONLY, pending production re-run · bundle sha256: a87246763088ca661e112e5b32ecd1fde4e80613e9cca9dc98f7afe19f2928c0. Details on the ARIN22 page.
All artifacts: github.com/saa-alliance/external-core-evidence, release v-r26. Verify: sha256sum -c SHA256SUMS.
Published research
Published research
The method is published under the founder’s name.
DRCET-1: Deterministic Risk-Compute Equivalence Test.
A Sampling-Resolution-Bounded Equivalence Benchmark for Institutional Risk-Compute Engines
Deterministic Risk-Compute Equivalence Testing
for Path-Dependent Functionals and Factor Attribution
Fail-Closed Release Gates for Numerical Operator Claims:
Certificates, Observed Order, and Spectral Domains in Risk Computation
Validating a Shared Deterministic Risk Kernel Across Two Peer Institutional Control Planes.
Backtest Evidence from Rolling-Window Validation and Cross-Product Implications
Low-Latency Stress Testing for AI-Integrated Macro-Financial Risk.
An Operator-Splitting Framework with Adaptive Error Calibration and Ontological Safety Controls
Adaptive Stress Testing via Operator Splitting.
A Multi-Operator Framework for Coupled Physical-Financial Risk Dynamics
Module-level methodology lives on each product page. Platform overview ->