This paper describes the reference implementation and exploratory evaluation of the MKMU ethical alignment framework, a companion systems contribution to the conceptual framework presented in 2. The implementation instantiates four Pali-form relational ethical primitives (metta, karuna, mudita, and upekkha) as culturally-conditioned LLM oracle scoring modules evaluated through a strict two-layer architecture. A key architectural contribution is this separation: a deterministic, LLM-free alignment gate (Layer 1) serves as the interface to the host processing system, while four culturally-conditioned LLM primitive scoring modules (Layer 2) operate as the ethical evaluation back end. The framework supports eight cultural conditioning profiles applied to each primitive oracle without altering the gate mathematics. We report results from an exploratory benchmark of 180 runs across three LLM models, four cultural profiles, and fifteen ethical scenarios. Results show inter-model agreement counts of 50 to 56 out of 60 scenario-profile cells across primitives, low cross-model score variance (mean standard deviation 0.020), and consistent classification of unambiguously harmful scenarios. Cultural profile conditioning produces observable verdict-level differences across profiles, providing exploratory empirical evidence that conditioning can produce measurable behavioral variation under fixed gate mathematics within this benchmark. All prompt inputs are reproducible via a full prompt and cultural profile version provenance chain recorded at evaluation time; LLM output nondeterminism is explicitly scoped out of this guarantee.
T. Bass (Sat,) studied this question.