BLADE-FINANCE Governance Node

Boundary-Layer Authority for Decision Enforcement, Financial-Sector Node - Reference Design and Simulation

A simulation-validated, software-enforced authority-arbitration reference architecture for financial-sector AI decision systems. BLADE-FINANCE places a hardware-anchored governance layer between automated transaction-decision models and consequential financial actions. It computes a four-tier authority disposition and appends every decision to a SHA-256 canonical-form evidence chain. It implements the eight-stage AUTHREX pipeline (VALIDATE, SATA, ADARA, MAIVA, HMAA, FLAME, ERAM, CARA) with a four-tier HMAA model (T3/T2/T1/T0), a population-state coordination model across account, device, payee, and IP-cluster history, and a retrospective stigmergic swarm-review module that recovers coordinated low-and-slow rings the per-transaction path clears. Tenth platform in the BLADE family and the first in the economic-security domain.

This is a reference design and simulation (TRL 3-4 simulation / TRL 2 hardware). No prototype has been built. All data are synthetic. The reported recall is an actionable-risk triage measure, not an empirical fraud-detection rate. The retrospective swarm review claims ensemble agreement and ring detection, not Byzantine fault tolerance. Per-record ECDSA P-256 signing and HSM key custody are design-specified, not exercised in the browser. The work has not been deployed in any financial institution. No federal endorsement is claimed.

Publication

DOI: 10.5281/zenodo.20374692
Author: Burak Oktenli · AUTHREX Systems; Georgetown University, M.P.S. Applied Intelligence
ORCID: 0009-0001-8573-1667
License: CC BY 4.0 · Version: v1.0 · May 2026

Policy and Standards Drivers

• U.S. Treasury Financial Services AI Risk Management Framework (FS AI RMF) - implementation framing for govern, map, measure, and manage functions applied at transaction-decision time.
• NIST AI Risk Management Framework 1.0 - foundation for the authority-tiering and evidence requirements.
• Executive Order 14179 - cited as implementation-framing policy context, not as endorsement.

Key Specifications

• Governance Plane: Xilinx Kria K26 SOM, hosting the AUTHREX pipeline and emitting ECDSA-signed per-stage audit entries
• Inference Plane: NVIDIA L4 Tensor Core GPU, hosting the transaction-decision model on a plane separate from governance
• Host Security Plane: Intel Xeon-D 1747NTE with Infineon SLB 9670 TPM 2.0 for measured boot
• Key Custody: YubiHSM 2 within a FIPS 140-2 Level 3 tamper-evident enclosure (epoxy potting, screw-continuity zeroization)
• Network: dual Intel X710-DA2 10GbE SFP+ for transactions; isolated 1GbE for management
• Enclosure: 1U rack-mount (19 inch EIA-310), redundant 600W 80-Plus Platinum PSUs, six 40mm N+1 fans
• Reference BOM: 36 components, 91 units, approximately $9,228

Eight-Stage AUTHREX Pipeline (Financial Instantiation)

• VALIDATE: Schema and canonical-form validation of the transaction record; malformed input rejected before scoring
• SATA: Input-integrity and authentication-trust scoring across channels; low-trust inputs excluded from the decision
• ADARA: Adversarial and deepfake / synthetic-identity signal assessment
• MAIVA: Multi-signal agreement aggregation across model, rules, and population-state evidence
• HMAA: Authority-tier arbitration (T3/T2/T1/T0) from upstream outputs and the population-state coordination score
• FLAME: Latency-bounded scheduling; fails closed to a more restrictive tier on budget exhaustion
• ERAM: Residual-risk gating before final disposition
• CARA: Recovery and manual-hold path; emits corrective audit entries

A retrospective stigmergic swarm-review module runs off the decision path, re-examining cleared traffic to recover coordinated rings that stay under per-transaction thresholds.

SHA-256 Evidence Chain

• Every stage output and final disposition serialized in canonical form and appended to a hash-chained, append-only audit ledger
• Per-record ECDSA P-256 signing and HSM key custody are design-specified at the hardware layer (not exercised in the browser simulator)
• Tamper-evidence is verifiable: a fixed seed reproduces a fixed ledger head hash and running checksum

Validation Scenarios (Simulator)

• A Nominal transaction - baseline legitimate activity - autonomous clearance
• B Deepfake authentication attack - synthetic-identity / liveness spoof pressure on VALIDATE and SATA
• C AI-agent coordinated attack - automated multi-account pressure on ADARA and MAIVA
• D High-value / risk signals - elevated-amount transactions exercising HMAA tier escalation
• MIX Realistic mixed stream - blended legitimate and adversarial traffic
• E Low-and-slow ring - coordinated ring the per-transaction path clears, recovered by the retrospective Swarm Review module

The engine is deterministic (mulberry32). Stepping the deterministic stream to 1,000 records at seed 42 fixes the audit-ledger head hash to 7d5aaab4... and the running checksum to 0x06acd6be. Triage confidence intervals are Wilson score intervals over a 2,000-trial run. The self-test suite reports nine of nine passing.

Standards Alignment

• U.S. Treasury Financial Services AI Risk Management Framework (FS AI RMF)
• NIST AI Risk Management Framework 1.0
• Executive Order 14179 (implementation-framing policy context)
• SHA-256 canonical-form serialization for the tamper-evident evidence chain

Related Work

• SATA: 10.5281/zenodo.18936251
• HMAA: 10.5281/zenodo.18861653
• CARA: 10.5281/zenodo.18917790
• ADARA: 10.5281/zenodo.19043924
• MAIVA: 10.5281/zenodo.19015517
• FLAME: 10.5281/zenodo.19015618
• BLADE-EDGE (defense): 10.5281/zenodo.19177472
• BLADE-AV (automotive): 10.5281/zenodo.19232130
• BLADE-MARITIME (maritime): 10.5281/zenodo.19246785
• BLADE-INFRA (critical infrastructure): 10.5281/zenodo.19277887
• BLADE-SPACE (orbital): 10.5281/zenodo.20183269
• BLADE-CUAS (counter-UAS): 10.5281/zenodo.20299604
• BLADE-AGENT-HSM (agentic AI): 10.5281/zenodo.20299821
• BLADE-SWARM (swarm autonomy): 10.5281/zenodo.20351198
• BLADE-INFRA-OT (IT/OT bridge): 10.5281/zenodo.20342067

Author

Burak Oktenli
AUTHREX Systems · Georgetown University, M.P.S. Applied Intelligence
ORCID: 0009-0001-8573-1667
Website: burakoktenli.com