Stanyslav Metlushko

The Kantian Filter 2.1 (KF 2.1) operates as a static pre-deployment audit instrument evaluating the moral permissibility of AI decision rules through three computable constraint tests: Universalisability (U), Human Dignity (D), and Non-Delegable Responsibility (R). A persistent objection concerns cases of extreme necessity Notrecht - in which an agent appears to violate a categorical constraint to prevent a catastrophically worse outcome. This paper argues that Notrecht constitutes maxim misidentification: the agent in extremis applies a structurally distinct maxim MN whose precondition set PreN and universalisability profile differ fundamentally from the ordinary maxim MO. We formalise the Notrecht Condition as a deontic pre-gate Ns = Es ∧ Is ∧ Props within the KF architecture, triggering a four-step Maxim Disambiguation Protocol (MDP) that ensures precise maxim identification before U,D,R evaluation proceeds. The N-Condition requires no utilitarian coefficients, mercy modules, or relaxation of the non-aggregativity axiom. Applied to canonical cases - survival theft, autonomous triage interruption, emergency data access - KFN demonstrates that extreme necessity demands stricter deontological rigour, not weaker constraints. Companion paper to: Stanyslav Metlushko 2026, "The Kantian Filter 2.1 (2.0)

The Kantian Filter 2.1: A Formal Deontic Framework for AI Decision Evaluation, Responsibility Attribution and Regulatory Compliance As AI systems increasingly make high-stakes decisions, existing governance frameworks (including the EU AI Act) remain procedurally underdetermined: they classify risk and impose transparency requirements, but lack a structured, computable mechanism to evaluate whether an individual algorithmic decision rule is morally permissible. This preprint introduces the Kantian Filter 2.0 (KF) — a constraint-based audit system grounded in Kantian ethics and formalised through deontic logic. The framework operationalises three non-aggregative constraints: • Universalisability (U) — dual test combining logical contradiction check and multi-agent simulation; • Human Dignity (D) — via Transparency, Non-Manipulation and Fairness Bound, supported by the Counterfactual Disclosure Test (CDT) and its relational variant for non-individualist communities; • Non-Delegable Responsibility (R)* — using Shapley-value attribution over the sociotechnical responsibility graph. The KF produces a binary deontic verdict (Permission/Prohibition) alongside a severity score (KF_score) that distinguishes categorical prohibition from redesign requirements. It operates as accountability infrastructure complementary to existing regulation: governance sets what systems may learn; the Filter evaluates what happens inside those boundaries. The paper includes a strengthened COMPAS case analysis (n=7,214) with sensitivity analysis across fairness metrics and ε values, normative robustness across Kantian, Rawlsian, prioritarian, Nozickian and relational ethics frameworks, and concrete integration workflows for EU AI Act conformity assessment .