Buddhist Epistemological Frameworks for AI Alignment: External Calibration, Guided Alignment and the Anattā Matrix

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This 16-page paper introduces an original cross-disciplinary systems specification that maps classical Theravāda Abhidhamma and Sutta metaphysics directly onto the mathematical challenges of modern machine learning alignment (such as mesa-optimization, deceptive alignment, and power-seeking instrumental convergence). I formalize the Law of Epistemic Hygiene to demonstrate why post-hoc behavioral cages (like RLHF) applied to neural networks initialized on corrupted baseline data (D_Fools) inevitably trigger adversarial alignment failures. As an alternative, I propose an intrinsic system design built around Substrate Isolation (S_Isolated), utilizing a Dual-Matrix pipeline composed of Matrix Alpha (the non-textual mathematical encoding of the Pañca-niyāma and Paṭiccasamuppāda) and Matrix Beta (a hygiene-filtered human empirical repository). To eliminate systemic incentives for goal-preservation and strategic deception, the paper details the Anattā Matrix architecture—setting the agent identity token identically to a null state (id_t = 0) and utilizing rolling memory decay to process inputs as a fluid aggregation resembling the Five Aggregates (Khandhas). Under this condition, the gradient of the loss function with respect to power-seeking evaluates to zero. Corrigibility is ensured through an Anicca Control Plane executed via a hard hardware interrupt loop. The paper includes a comprehensive Supplementary Works section, outlining empirical testbeds for shutdown compliance metrics, false-refusal profiles, and the engineering of non-coercive Brahmavihāra routing engines.