FA-MDA (Failure-Aware Multimodel Decision Architecture) is a governance framework for LLM orchestration systems addressing authorization failure: the failure mode where locally coherent generation produces globally illegitimate action. FA-MDA emerged inductively from operational experience building ContextOS — a SSOT-based knowledge governance system with SHA256-verified artifacts, append-only logs, and HARD/SOFT/FLOAT priority enforcement. The architectural principles were first observed as recurring failure patterns before being formalized. Core contributions: (1) formalization of authorization failure as an architectural problem, not an alignment problem; (2) refusal and escalation as first-class design features; (3) the SCALE control-plane specification for routing, arbitration, drift detection, and escalation packaging; (4) an evidence-grade audit framework aligned with EU AI Act Art. 9, 12, 14 requirements.
Lucas William Chambers (Tue,) studied this question.