ATLAS: Artifact Generation Through Layered Constraints and LLM x MDE Synergy

ATLAS unifies Large Language Models with Model-Driven Engineering to generate regulator-ready artifacts and machine-checkable evidence for safety- and compliance-critical domains. ATLAS integrates three pillars: a Unified Meta-Model (UMM) reconciles heterogeneous schemas and regulatory text into a single semantic space; an Integrated Constraint Model (ICM) extends our prior Dual-Stage(S2D2) extraction logic to compile layered requirements into deterministic generation-time automata (Layer~1) and post-generation validators (Layer~2); and Constraint-Guided Verifiable Generation (CVG) applies these through two-layer enforcement -- Layer~1 structural constraints drive prefix-safe decoding while Layer~2 semantic/logical validation produces machine-checkable certificates. When violations occur, ATLAS performs audit-guided repair and records generation traces for compliance review. We evaluate ATLAS in automotive software engineering (AUTOSAR) and cross-border legal jurisdiction (Brussels~I~bis). ATLAS produces structurally valid, auditable artifacts that integrate with existing tooling and substantially reduce manual remediation effort, validating a graduated automation paradigm that automates routine construction while empowering experts to resolve complex semantic ambiguities through machine-checkable evidence.