A runnable demonstration that data-agent accuracy comes from governance context, not the agent loop — measured with a four-level ablation ladder on the BIRD benchmark. - View it on GitHub