A good explanation suggests an action that could have changed the outcome. For spatio‑temporal GNNs, that means minimal feature/edge changes that flip the decision and are physically plausible.

“Actionable explanations are the shortest paths to correction.”

Research as craft: iterate in small, complete loops; maintain a living notebook; test hypotheses quickly; automate once; document always.

• Bias toward small, demo‑able artifacts.
• Name and version everything.
• Write docs as if a stranger will thank you.