Authority–Inference Separation in Agentic Finance: First-Line Control, Blockchain Enforcement, and Replayable Assurance

Battle Count: 51

The paper addresses governance and control architecture for agentic AI in finance rather than trading strategies, signal generation, or portfolio construction. While it touches on investment operations (order routing, pre-trade compliance, best execution) and treasury (FX, liquidity), these are framed as control-plane examples rather than quantitative trading methods. The relevance is indirect: any quantitative trading system deploying AI agents for order execution, rebalancing, or data procurement would need such authority-inference separation, but the paper does not contribute to alpha generation, execution optimization, or risk modeling per se.

The full AIS architecture involves five planes (mandate/identity, adaptive inference, deterministic authority/control, execution/settlement, evidence/handoff), a meta-control layer, blockchain smart-contract verifiers, EIP-3009 authorization, nonce consumption, canonical intent hashing, signed tokens, reason-code services, delivery reconciliation, accounting mapping, and replayable evidence bundles. Cross-domain instantiation across treasury, trade finance, lending, and investment operations requires domain-specific policy rules, approval matrices, and evidence schemas. The deterministic prototype is simpler (48 fixtures, 3 configurations), but production deployment would require integration with existing IAM, mandate registries, policy engines, blockchain infrastructure, accounting systems, and three-lines governance. The encodable-attestable-human taxonomy adds classification complexity.

The paper explicitly preserves v2 code, machine-readable fixtures, event records, results, summary, and protocol separately from v1 in a replication directory. A claim-boundary checklist (Appendix B) mandates preserving source URLs, retrieval times, raw payloads, derived transformations, code/fixture versions, and checksums. The deterministic prototype uses fixed fixtures with no live model calls, making results fully reproducible. However, the public-ledger sample is time-sensitive (captured 9 August 2026) and non-random, limiting exact reproduction of that component. No explicit GitHub URL is provided in the extract.