How it works

From sensor nodes to a single audited verdict.

A decentralised, event-driven pipeline with a 22-agent orchestration layer. Heterogeneous signals enter through Sensor Nodes, cross a high-throughput event bus, and arrive at the ARIN Council, where calibration-aware confidence-and-reliability weighting produces one fully audited verdict. Six domain modules plus the KOKON control-plane module — each bounded, independently deployable, and governed under one release discipline. The ARIN22 deterministic kernel is positioned as an independent challenger and governance layer — embeddable beside Aladdin, Bloomberg, Murex, MSCI, Calypso, AxiomSL, and internal bank Monte Carlo engines, not a replacement for any of them.

Pre-client institutional diligence-ready — H100 validation pack closed end-to-end (manifest-scoped H100 pack: 58/58 data-room + 53/53 closure artifacts, 0 missing); external pilot and regulatory sign-off pending; not enterprise production-approved.

Commercial positioning

Independent challenger and governance layer.

SAA Risk Analyzer sits beside existing institutional risk systems — Aladdin, Bloomberg, Murex, MSCI, Calypso, AxiomSL, and internal bank Monte Carlo engines. The customer connects existing risk outputs; SAA independently recomputes VaR / CVaR / ES / scenario risk on the ARIN22 deterministic kernel.

Differences surface as PASS / WARN / FAIL / INCONCLUSIVE evidence. The customer receives an audit-grade challenger evidence pack: snapshot hash, model version, policy version, and cannot-claim boundary — committee-readable, regulator-ready.

Product category: sealed institutional risk-compute workstation, analog in delivery model to Bloomberg Terminal or Refinitiv Eikon. Not a data API, not developer middleware, not an exchange execution engine.

AI architecture

Deterministic numerical analysis. AI as narrative layer only.

Every numerical output — fair value, scenarios, VaR, CVaR, ES, risk metrics, BUY / HOLD / SELL research verdicts (non-advisory) — is deterministically computed by the math kernel. Rule-based logic produces all decisions. The LLM is restricted to narrative prose with deterministic template fallback. If NVIDIA NIM is unavailable, the report still renders fully from templates with computed numbers.

Numerical outputs0% AI. No AI-generated fair values, VaR, CVaR, ES, scenario outputs, or risk calculations.

Research verdicts0% AI. BUY / HOLD / SELL and PASS / WARN / FAIL / INCONCLUSIVE are rule-based.

Graphs0% AI. Charts are rendered from computed series.

Narrative prose~10–15% AI. AI is prose polish only; the LLM layer can be disabled without losing core product functionality.

Math first · Agents second · Governor always · Audit forever.

Pipeline schematic

One flow: route, ingest, compute, govern, deliver, audit.

Scientific Analytics Alliance

ARIN pipeline — live schematic
ARIN22 deterministic — 0% AI in numbers

Discover & Route

The operator selects an analytical domain; typed contracts route the request to the module that owns it.

Ingest & Normalize

Heterogeneous signals are normalized, hashed, timestamped, and published to the event bus.

Compute — ARIN22 Kernel

VaR, CVaR, ES, scenario risk, stress, and cascade math are deterministic. LLMs do not calculate numbers.

Orchestrate — ARIN Council

The council deliberates on the deterministic surface with calibration-aware reliability weighting.

Govern — Refusal Gates

Named gates hold release when substrate, math reconciliation, or evidence quality is insufficient.

Stitch & Deliver

Approved output is stitched with lineage and dual-published through API and WebSocket surfaces.

Audit Forever

Historical verdicts can be rerun through input, output, hash, policy, and evidence trace.

The pipeline · seven stages

01Discover & Route

The operator selects an analytical domain across the six domain modules and KOKON control-plane module; the request is routed over typed contracts to the module that owns it. Each module runs on its own domain with isolated infrastructure — no shared single point of failure. Domain-bounded routing prunes the agent graph to only the agents relevant to the query.

02Ingest & Normalize

Heterogeneous signals — market, geopolitical, on-chain, narrative — enter through Sensor Nodes and are normalised onto typed contracts, then published to a high-throughput event bus (HMAC-signed delivery, idempotent, dead-letter-protected). Every input is hashed and timestamped on arrival.

03Compute

Reproducible risk math — VaR / CVaR, stress, cascade — is computed by the ARIN22 deterministic kernel, never by an LLM. Kernel correctness is CPU-bound: ~99 µs single kernel call at D=4 on commodity ARM CPU, substrate-portable with no GPU lock-in. Scaling is validated on the D=200 high-dimensional lane: 200 portfolios × 28 scenarios × 4 horizons × 10M paths per case, 1.344 trillion total GPU+CPU paths, median GPU wall 30.014s vs CPU 430.596s (14.35× speedup). The 8×H100 lane is layered on top for batch throughput at scale; the two lanes are never collapsed into one number. Operator Contract v1 defines class-routed operators across diffusion, drift, jump, regime, and tail dimensions; public surface is input → output → hash → lineage, with implementation details released only under signed NDA to bank model-risk reviewers.

04Orchestrate · ARIN Council

The 22-agent council deliberates on the deterministic surface under NeMo Guardrails; calibration-aware confidence-and-reliability weighting resolves consensus, and diverging agents lose voting weight automatically. Before anything publishes, the Governor stack runs 14 named constitutional refusal gates, including SYS.GOV.01.SUBSTRATE.NO_RELEASE, SYS.GOV.01.MODEL_AGREEMENT.NO_RELEASE, SYS.GOV.01.MATH_RECONCILIATION.NO_RELEASE, and SYS.GOV.01.BAYESIAN_EVIDENCE_FUSION.NO_RELEASE. Each gate blocks the verdict until cleared; the system refuses to publish rather than degrade silently.

05Stitch & Deliver

A single explainable verdict is stitched with its full lineage and dual-published — synchronous API response plus a WebSocket push — each carrying the cryptographic decision spine. A Zero-Outcome Freeze holds any verdict that cannot resolve its evidence; nothing degrades silently.

06Resilience

Graceful degradation across the substrate: if a data source, agent, or compute lane fails, the pipeline holds correctness over availability and records the degradation in the audit trail. Webhook-isolated topology; multi-domain isolation; deterministic replay lets any historical verdict be re-executed bit-for-bit.

07Audit Forever

Historical verdicts can be rerun through input, output, hash, policy, and evidence trace. The audit trail keeps release decisions replayable, source-weighted, and regulator-readable without exposing protected kernel construction.

Live access

Public access is limited to verification and marketing surfaces. Full module access, replay artifacts, and bank model-risk packets are private and released under NDA. Audit-grade lineage is published in parallel at /arin22-demo.

Verification surface/arin22-demo — audit-grade lineage, determinism proof, evidence matrix.

Platform overview/platform — the seven modules & architecture.

Team & method/team · /methodology — founder profile and published method.

EngineeringGitHub org · ARIN architecture (PDF).

Execution-venue phase mapPhase 0 GUARDED_INTERNAL · Phase 1 DESIGN-PARTNER READY · Phase 2 REVIEW ONLY · Phase 3 REVALIDATION REQUIRED · Phase 4 DEFERRED.

Validation evidence matrix

Observed, not asserted.

Enterprise Wave v2 — canonical · 2026-05

8.8 B

paths / backend (GPU grid + CPU mirror)

8,800

Enterprise Wave cases, 0 execution failures

0.035%

CVaR 99.9 vs a 10M-path MC reference

0.000 bp

CRN 99.9 parity Δmax on 100 worst-tail cases × 100M paths / backend

1.344 T

D=200 high-dimensional GPU+CPU paths

14.35×

D=200 GPU wall 30.014s vs CPU 430.596s

1 / 1,000

one hash across 1,000 fresh-process repeats

58 / 58

data-room artifacts (+ 53/53 closure)

Every figure is ladder-able to evidence at /arin22-demo.

Substrate breadth · context (not the headline)

V1 substrate lineage1.06 T path-equivalents · 1,043,030 cases — composition disclosed in the V1 lineage pack: 8-Wave full GPU census 705,096 cases plus parity, deep-tail, tick-replay, feed-chaos, shadow-ledger, and closure lanes; substrate breadth, superseded as the headline by Enterprise Wave v2.

Unified Public H100 Run · 2026-Q2705.096 B paths · 705,096 cases — public-portfolio breadth run; context, not the canonical accuracy anchor.

D=200 high-dimensional lane · 2026-05-31200 portfolios × 28 scenarios × 30/90/180/360D × 10M paths per case. N=3 GPU + N=3 CPU mirror repeats · CVaR99.9 p95 max-across repeats 0.281% · CVaR99.9 max 0.900% · EVT-CVaR99.9 max 0.671% · status: internal high-dimensional proof; client full-holdings D=200 production validation remains external-data-gated.

Stress Propagation Engine62,500 historical cascade cases across seven domain profiles · F1 = 0.889 cross-domain average · F1 = 0.914 climate / physical / utility / production · zero negative-control false positives · sealed verdict bundles under arin.report_audit_trace.v1.

Governance controls · Governor refusal stack

Every council pass is gated before publication.

Surface 01 · 5 Decision-Integrity InvariantsI-01 calibration sanity · I-02 reliability stationarity · I-03 consensus coherence · I-04 dissent representation · I-05 evidence-source completeness. A verdict must pass all five.

Surface 02 · 7-Rule Math-to-Narrative AuditR-01…R-07 bind the explanation to the numerical surface; any drift between narrative and the deterministic state fails the audit and the verdict is held.

Surface 03 · Meta-Decision GovernorFour terminal states — APPROVED / WATCH / REVISION / BLOCKED (e.g. WATCH_SUPPRESSED on EVT 5Y/10Y). Curvature-heavy cases — where proxy methods carry unbounded tail error — route to full revaluation; no silent degradation, and the route is hash-logged.

14 named refusal gatesThe 5 + 7 + Governor framing maps to a 14-gate constitutional refusal stack. Selected named rules: SYS.GOV.01.SUBSTRATE.NO_RELEASE, SYS.GOV.01.MODEL_AGREEMENT.NO_RELEASE, SYS.GOV.01.MATH_RECONCILIATION.NO_RELEASE, SYS.GOV.01.BAYESIAN_EVIDENCE_FUSION.NO_RELEASE. Each gate blocks the verdict until cleared.

Output surface · what MRM receives

What is handed to the customer’s MRM function on every approved verdict.

Decision evidence spine5-layer cryptographic spine, SHA-256 chained — input snapshot, kernel & model versions, governor decisions, calibration state, output hash.

SR 26-2 post-run checkerMulti-pillar gate run after every approved verdict; aligned with Fed / OCC / FDIC interagency Model Risk Management guidance, EU AI Act Article 12, and DORA (EU 2022/2554) where applicable.

Shadow Ledger · backtest harnessArmed; REALIZED_BACKTEST_PASS is forbidden before 250 matched observations — a gate, not a policy.

Reproducibility snapshotBit-for-bit replay store; any historical analysis re-executes to the same hash.

Trace export · regulator-safeFull lineage export with no kernel disclosure — public proof, secret construction.

Determinism proof1 hash across 1,000 fresh-process repeats.

Pre-trade gate · three surfacesMixed gate (24M lane): 24M orders, 88.83% ALLOW / 11.00% REVIEW / 0.16% BLOCK, p99 1.064 ms. Mixed gate (50M lane): ALLOW 9,999,675 / REVIEW 40,000,124 / BLOCK 201, p99 band 0.85–0.93 ms, p999 band 1.8–2.5 ms. Stress gate (50M lane): BLOCK 50,000,000 by zero false-ALLOW construction, p99 0.84–0.87 ms, p999 up to 2.409 ms.

Latency disclaimerInternal risk-service compute latency only — not STAC-T1, not wire-to-wire, not NIC-to-order-egress, not broker / order-routing latency.

DRCET-1 protocolOpen Equivalence-Testing Protocol submitted to SSRN on June 7, 2026 (Abstract ID 6893318, preliminary upload). Sampling-resolution-bounded equivalence benchmark with Berger-Hsu interval inclusion against Monte Carlo bootstrap confidence intervals at Bonferroni-corrected 99.75% (k = 20), with mandatory WATCH disclosure rather than vendor overclaim.

Real-name attributionRecurring tail contributors recorded by frequency — observation, not forecast.

Live lineage dashboard/arin22-demo — every claim ladder-able to its evidence.

Evidence substrate

The council does not reason in a vacuum.

Hydra Ontology114 M entity nodes · 336 M relationships · 9.1 M RAG chunks (hash-pinned audit graph), plus ~500 M operational embedding (runtime H100 similarity routing, not the audit substrate). Every retrieved fact carries source, timestamp, and evidence weight; the 7-rule audit refuses any narrative claim lacking a corresponding subgraph.

5-LLM inference sovereigntyvLLM (local H100) → NVIDIA NIM → frontier reasoning → alternate reasoning → local quantised fallback. NVIDIA-first, NeMo-guardrailed, DORA-aligned — no vendor lock-in. All LLM lanes are restricted to the narrative-prose layer; numerical computation is deterministic kernel only.

Regulatory & compliance frame

Frameworks the pipeline is designed against — alignment, honestly tiered (see /security for per-framework status):

Model riskFed / OCC / FDIC SR 26-2 (interagency Model Risk Management guidance, Apr 2026; supersedes SR 11-7 / OCC 2011-12) · EU AI Act Article 12.

Market & capitalBasel III IMA · FRTB-IMA · BCBS · Pillar 3 · IFRS 9.

InsuranceSolvency II · IFRS 17.

Operational resilienceDORA (EU 2022/2554).

Security & privacySOC 2 Type II (audit window Q3 2026) · ISO 27001 · NIST CSF 2.0 · CCPA / CPRA · GDPR / UK GDPR.

External MRM ownership remains customer-side; the customer’s own book is the validation target. Full security & compliance →

Phase Map · honest roadmap

Clear states, no binding production promotion before closure.

Phase 0 · Verdict Integrity + SSOTGUARDED_INTERNAL. readiness check PASS · live export smoke PASS · stale PDF export fail-closed.

Phase 1 · Challenger / Regulatory WedgeDESIGN-PARTNER READY. PASS/WARN review export fixtures, live local API export, incumbent import bridge for Calypso VaR/CVaR, AxiomSL/FRTB ES, internal bank MC, challenger audit-ledger drilldown, partner payload mode, pilot bundle, dry-run runbook, and incumbent risk import contract.

Phase 2 · Pre-Trade Risk GovernorREVIEW ONLY. Typed market-data contract, six-order replay fixture (ALLOW 1 / REVIEW 3 / BLOCK 2), CPU mirror smoke, latency bands. No production ALLOW emission.

Phase 3 · Risk-per-tick / Auction risk-deltaREVALIDATION REQUIRED. Current-kernel smoke started: 9 portfolio-tick updates + 2 auction/cross cases. Load/parity: 216 processed updates + 4 auction/cross cases, p999 75.272ms local dev wall-clock, stale-input fail-closed proof, arithmetic/decision mirror PASS. Full promotion still requires 60M+ replay/load scale and true CPU performance mirror/parity.

Phase 4 · Surveillance risk-consequence overlayDEFERRED_RISK_CONSEQUENCE_ONLY. Typed external-event contract prepared. ARIN22 may compute portfolio risk consequence after a client/vendor/exchange event is supplied. Detection and enforcement remain out of scope; no SMARTS replacement claim.

Engagement Models

Pilot, module, enterprise, sovereign, or embedded kernel.

Pricing is disclosed under signed engagement. Each path includes a scope boundary, evidence pack, and cannot-claim matrix.

Paid Pilot · 4–8 weeksOne portfolio or one entity stress assessment. Challenger report, divergence analysis, evidence pack.

Module License · annualSingle domain: sovereign / city risk, insurance, banking, infrastructure, real estate, or asset management. Web UI + API + reports.

Enterprise License · annualMulti-domain ARIN + KOKON + Risk Analyzer + API + monitoring for multiple departments.

Sovereign / On-Prem / Air-GappedOn-prem deployment, private graph / RAG, custom integrations, and strict data-policy isolation.

Embedded Kernel / APIARIN22 as computational layer for existing risk systems; an accelerator for stress / tail / cascade analytics, not a replacement for incumbent platforms.

Walkthrough

A line-by-line review of the pipeline, governance, and evidence.

A 60-minute session — screen-share, founder Q&A — covering the seven stages, the output surface, the three governance surfaces, and the evidence matrix.

Request a pipeline & governance walkthrough

Pre-client diligence-ready with a strong public-portfolio H100 validation package, not yet externally production-approved; embeddable inside existing workflows, not a replacement for Aladdin / Bloomberg / Murex / MSCI.