Type: article
Abstract: Repeated delegation of operational decisions to AI-enabled systems raises architectural challenges around context, strategic intent, validation discipline, effect attribution, and reusable experience across cycles. This paper presents PDEL-DRK, an ablation-testable architecture with five typed modules: contextual perception, strategy-tactic decision, execution validation, effect attribution, and decision relation knowledge. We instantiate it in a synthetic hierarchical e-commerce seller-operation testbed and evaluate it over a 144-week paired ablation panel with an external contextual-bandit reference. Five of six pre-specified commitment ablations significantly reduce cumulative operating value after Bonferroni correction. Contextual perception is strongest; strategy-tactic separation and tactical relation retrieval form a second tier; execution validation and strategic retrieval are smaller. Effect attribution is not detectable under the pre-specified drift magnitudes, although an exploratory drift sweep suggests a boundary-sensitive role at the highest tested regime. The paper contributes nascent design theory by showing how long-horizon reliability decomposes into inspectable, typed, ablation-testable commitments.
Authors: Kim, Hoeyoung, Jang, Yunhwan
Language: English
Subfield: Management Information Systems
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7222988817