Marginal is an open-source governance layer for AI coding agents. It starts in Shadow Mode, collects verifiable runtime evidence, and only earns limited enforcement authority when the evidence supports it.