The Myrmidon Family: A Universal Pattern for Epistemic Agent Correctness — Formal Specification, Multi-Domain Validation, and Cross-Instance Analysis

Autonomous agents deployed in high-stakes domains fail not because their reasoning engines are flawed, but because they act on beliefs that are epistemically invalid — temporally stale, semantically contradicted, or spuriously corroborated. This failure is domain-agnostic: it occurs in financial fraud detection, adversarial simulation, UAV swarm operations, and clinical decision support with equal structural regularity. Existing agent frameworks, both LLM-centric (AutoGen, CrewAI, LangGraph) and classical symbolic (Jason, Jadex), provide no structural mechanism to prevent it. This paper presents the Myrmidon Family: a universal architectural pattern and four validated domain instantiations that resolve this failure by construction. The Myrmidon pattern mandates five structural layers in every instance — typed sensor, EVR epistemic gate, BDI-HTN cognitive engine, personality vector, and Myrmidon-OP coalition coordinator — and nine structural invariants that together guarantee epistemic correctness at any processing volume. Four instantiations span orthogonal domains: Myrmidon-Ω for AML and fraud detection (validated in production), Myrmidon-A for adversarial AML red-team simulation (with a formal proof of the Theorem of Absent Mens Rea), Myrmidon-Δ for autonomous UAV swarm operations (with a two-tier AQUILES cognitive-reactive architecture), and Myrmidon-Φ for clinical belief management. The central theoretical result is the Cross-Instance Universality Theorem: the Myrmidon pattern's epistemic safety guarantees are preserved under instantiation to any domain, with N* = ∞ in all four cases. Cross-instance analysis identifies three universal invariants (epistemic gate, deterministic cognition, structural class-L exclusion), two domain-specific axes (EVR calibration, personality archetype), and one substrate property (AQUILES derivation) that together explain why the pattern works across domains that share no domain knowledge. We compare the Myrmidon family against six competing systems (Project Maven, Gospel, AutoGen, CrewAI, LangGraph, Jason) across nine architectural dimensions, and show that the family is the only framework that simultaneously satisfies epistemic correctness, scale invariance, formal auditability, and structural human authority over irreversible actions.