Arbitration-Driven Development: A Governance Framework for Trustworthy AI-Assisted Software Engineering

As large language models (LLMs) become widely adopted in software engineering, the primary bottleneck has shifted from code generation to governance and assurance. Existing methodologies (e.g., TDD, ODD) mainly describe how to build, but often leave unresolved the question of how to trust what has been built - especially when AI can produce large code changes faster than humans can reliably review. This paper proposes Arbitration-Driven Development (ADD), a model-agnostic governance framework that redefines the engineer's role from author to arbiter. ADD is grounded in a "conservation of accountability" premise: regardless of how much work AI performs, accountability cannot be automated away and must ultimately be assumed by humans. To operationalize this premise, ADD introduces (1) risk-proportional governance via L1-L4 grading, (2) a Contract Adversarial Protocol (CAP) that engineers conflict before code is produced, and (3) a state-machine-based authority boundary centered on sealing/unsealing. We present a reference implementation in the Progee system and a narrative case study (points redemption) to illustrate how ADD surfaces ambiguity early, concentrates human attention on high-impact decisions, and establishes a clear audit trail for responsibility.