Abstract
<title>Abstract</title> <p>In omnichannel retail, customer data quality and unification architecture encompasses the systems and controls that turn fragmented customer observations into trusted, unified profiles, yet rigorous evaluation is constrained by limited access to shareable records with known identity, defect, data-lineage, and survivorship ground truth. This paper presents a synthetic Customer 360 benchmark and generation method for controlled evaluation of identity resolution, data-quality transformations, and golden-record survivorship rules. The benchmark models persons, households, accounts, contact points, addresses, source observations, and transaction or order context across four profiles: ecommerce, loyalty or mobile, physical-store transaction, and service or fulfillment. It generates structural duplicates and eight auditable defect, ambiguity, and conflict mechanisms while preserving separate ground truth for duplicate clusters, transformation lineage, and expected outputs under verification-first, source-authority-first, recency-first, and mixed-conditional regimes. The study evaluates the benchmark through 335 predeclared runs organized into identity-difficulty, survivorship-conflict, single-factor-sensitivity, and scale experiments. Three transparent identity baselines are reported with pairwise and B-cubed measures; survivorship is assessed through applicability, evaluability, union and pairwise disagreement, conditional disagreement rate, and population-level burden. All 335 confirmatory runs passed validation. In 40 paired seeds, the easy identity condition exceeded the hard condition for weighted-similarity pairwise F1 at the primary 0.85 threshold, with a median paired difference of 0.5846 and a Holm-adjusted one-sided p value of 1.82 × 10⁻¹². The direction held at predeclared thresholds of 0.80, 0.85, and 0.90. High-conflict survivorship conditions exceeded low-conflict conditions in all 40 paired seeds, with a median conditional-disagreement increase of 0.0577 and the same adjusted p value. These findings demonstrate reproducible condition separation and auditable ground truth under the disclosed synthetic design. They do not establish real-retailer prevalence, operational effectiveness, production scalability, or universal superiority of any resolution or survivorship rule.</p>