The European Union Aviation Safety Agency (EASA) established a rigorous phased framework for AI trustworthiness in aviation. We explore the architectural path to certify Level 1B and Level 2 agentic co-pilots.
Traditional DO-178C software development assumes deterministic source code coverage. Machine learning models require Learning Assurance under EASA guidelines, proving training data completeness, distribution coverage, and robustness against out-of-distribution inputs.
Clear authority boundaries are essential. Our implementation maintains the human pilot as final authority (Level 1B/2), utilizing explicit confirmation gates for irreversible control actions while automating tactical monitoring.