Coherence as First Principle: Deriving AI Safety Requirements from Axiomatic Foundations

AI safety frameworks typically assert their requirements. Reinforcement learn-ing asserts that human preferences can guide model behavior. Constitutional AIasserts a set of governing principles. Alignment theory asserts that AI goals shouldmatch human values. None derive its requirements from a foundation that survivesphilosophical scrutiny. Who can answer the question why these requirements andnot others?We present “Contextual Epistemic Dynamics for Amplified Humanity” (CED-AH), a framework that derives safety requirements from a single first principle:coherence. Through a transcendental argument, we show that the rejection ofcoherence as an epistemic criterion is self-defeating. Any attempt to argue againstcoherence must itself be coherent to have argumentative force. This is not a novelphilosophical invention but effectively distills what evaluation already presupposes.From this principle of coherence, through six defended commitments and fourbridging principles, we derive necessary conditions for safe computational systems.These include: the instrument classification of computational artifacts, the require-ment to amplify rather than substitute for human cognitive authority, and five spe-cific design imperatives—Authority Retention, Capability Growth, Transparency,Calibrated Honesty, and Reversibility.The derivation necessitates a reframe from “Artificial Intelligence” to “AmplifiedHumanity”—not as rhetorical preference but as a consequence of the ontologicaland ethical analysis. Computational systems are human-dependent artifacts; theirproper function is the amplification of human capacity, not the replacement ofhuman agency.We identify the framework’s load-bearing commitments and their contestableconditions, and indicate where the derivation establishes necessity and where it1 establishes well-defended contingency.The coherence derivation produces two foundational safety commitments. Thispaper develops the first: the instrument thesis, which classifies computational sys-tems as instruments for human flourishing. The second—the accountability prin-ciple, which requires that deploying entities bear responsibility for their deployedsystems’ effects—is derived from the same coherence criterion and developed fullyin a companion paper (Doane, 2026c). Together, these two commitments form theframework’s structural foundation: the instrument thesis defines what computa-tional systems are; the accountability principle defines who answers for what theydo.The design implications of these derivations, including measurement criteriaand domain applications, are developed in a companion paper (Doane, 2026b).The framework’s failure modes, falsification conditions, and limits are analyzed ina third (Doane, 2026a).