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.

翻译：ATLAS将大语言模型与模型驱动工程相统一，为安全关键和合规关键领域生成符合监管要求的制品及机器可验证证据。ATLAS集成三大支柱：统一元模型（UMM）将异构模式与监管文本协调至统一语义空间；集成约束模型（ICM）扩展了我们先前的双阶段（S2D2）提取逻辑，将分层需求编译为确定性生成时自动机（第1层）与生成后验证器（第2层）；约束引导可验证生成（CVG）通过双层机制实施这些约束——第1层结构约束驱动前缀安全解码，而第2层语义/逻辑验证生成机器可验证凭证。当出现违规时，ATLAS执行审计引导的修复并记录生成轨迹以供合规审查。我们在汽车软件工程（AUTOSAR）和跨境法律管辖（Brussels I bis）场景中对ATLAS进行评估。ATLAS生成结构有效、可审计的制品，这些制品能与现有工具链集成，并显著减少人工修复工作量，从而验证了一种渐进式自动化范式——该范式自动化常规构建过程，同时通过机器可验证证据赋能专家解决复杂的语义歧义问题。