AEGIS: An Auditable Evidence-Governed Interface for Cost-Aware AI Harness Selection in Finance

Xiaozhen Wang () and Francois Buet-Golfouse ()
Additional contact information
Xiaozhen Wang: CEREMADE - CEntre de REcherches en MAthématiques de la DEcision - Université Paris Dauphine-PSL - PSL - Université Paris Sciences et Lettres - CNRS - Centre National de la Recherche Scientifique
Francois Buet-Golfouse: AIML Global Markets, Barclays

Working Papers from HAL

Abstract: In bank onboarding, an accurate model output may remain non-executable until required evidence and approvals are in place. Obtaining them consumes money, compute, and human capacity and adds latency, yet may unlock a higher-value action. We introduce the Auditable Evidence-Governed Interface for Selection (AEGIS), a cost-aware framework that selects among versioned, pre-commit-mediated harnesses by conditioning jointly on task information, policy, evidence, budget, deadline, and runtime health. At fixed resource and risk prices under a stationary cell law, our frontier theorem identifies the support gap \(D(\theta)\) as the per-decision value of governance information. This value vanishes exactly when each task fibre has a common cost-adjusted optimum; otherwise, task-only routing incurs \(T D(\theta)\) expected structural pseudo-regret. LP duality yields auditable supporting prices for resources and risk; independent read-once evidence packages admit an exact assurance-option rule. Five repeated outer-CV evaluations of 1,000 borrowers show that governance-conditioned routing lowers normalised loss .099 versus task-conditioned routing (borrower-by-repeat 95\% sensitivity interval [.081,.116]) and .090 versus a training-selected fixed policy [.066,.111]; operations loss falls 33.2\% for .018 higher business loss. Across 629 matched cells in four tool-use suites, AEGIS supplies nine of ten cell-weighted and 14 of 15 equal-suite observed nondominated safety--risk--latency--call points. At one resource price, descriptive held-out-goal estimates show 7.31 points higher safe completion, .79 points lower attack success, 1.788 fewer seconds, and .113 fewer calls than suite-fixed. Signed evidence certifies lineage and specified checks, not factual correctness.

Keywords: Agentic AI; AI governance; Financial model risk; Cost-aware routing; Runtime assurance; Evidence acquisition (search for similar items in EconPapers)
Date: 2026-09-02
Note: View the original document on HAL open archive server: https://hal.science/hal-05735024v1
References: Add references at CitEc
Citations:

Downloads: (external link)
https://hal.science/hal-05735024v1/document (application/pdf)

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:hal:wpaper:hal-05735024

Access Statistics for this paper

More papers in Working Papers from HAL
Bibliographic data for series maintained by CCSD ().