This document presents the Understanding-Aligned Intelligence Framework (UAIF), an architectural proposal and research roadmap for achieving provably safe artificial intelligence through genuine cognitive understanding rather than mere behavioral compliance. UAIF integrates formal verification methods from Martin-Löf Type Theory with constitutional principles derived from international human rights law (UDHR). The framework implements a Proof-Carrying Code (PCC) architecture that separates decidable verification (O(n) time) from undecidable proof generation, explicitly acknowledging that automated proof generation for ethical constraints remains an open Grand Challenge. Key contributions include:- Five-layer architecture (Constitutional Principles → Understanding Engine → Motivational Constraints → Rule Validation Gateway → SI Implementation)- Epistemic Adequacy / Motivational Conformance (EA/MC) split addressing the "psychopath problem"- Proportionality calculus based on Alexy's Weight Formula for rights conflict resolution- Comprehensive threat model covering external attackers through mesa-optimizers- Explicit limitations including Frozen Robot Problem and Reality Gap UAIF is positioned as a rigorous foundation for AI safety research rather than a deployment-ready specification. The framework builds on the Cognitive Understanding Architecture (CUA) and provides a formal basis for future work in provably aligned AI systems. Version 1.0.0 includes 7 architectural diagrams, complete references, and detailed analysis of open research challenges.