The software moat only really holds if you can explain regressions or drifts across fleet updates. How are you tracing what changed when compression a model, you manage runtime versions, and/or have variable on-robot conditions that all move at once?