Type: article
Abstract: SCORSE (Structured Contextual Orchestration and Reward Shaping Engine) is a computational design artifact for advisor-in-the-loop financial recommendation drafting. It combines parser-visible structured initialization with transparent heuristics (H), a cross-model semantic/audit judge (J), scenario performance testing (P), targeted repair, and noncompensatory gates. In a fixed-seed benchmark of 50 synthetic clients, iterative modes improve machine-measured semantic reasoning and the Reasoning Quality Index relative to scaffold-only initialization. SCORSE-HJP raises the machine pass/review-ready rate from 18% to 44%, although a simpler judge-only loop is competitive, narrowing the claim from full-stack necessity to value beyond structure plus governable evaluator separation. Sensitivity analysis shows that continuous improvements are stable under reasonable reweighting, while binary clearance depends on calibration. Boundary analysis identifies remaining failures in personalization, overclaim suppression, semantics, and feasibility. The contribution is a prototype-level design architecture for inspectable, stress-tested recommendation drafts, not evidence of practitioner-certified advice.
Source: AIS Electronic Library (AISeL)
Institution: University of South Florida
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222992440
