Temporal safety alignment
Learned from hindsight — no extra labels.
Hindsight relabeling traces each near-collision back to the frames that led to it, teaching the policy an early, graded response to danger — at matched speed, not by slowing down. But fixing danger naively damages frames that were already safe; EWC is what lets us get both. Step through it.
Implementation: hindsight relabeling sorts frames into three tiers — normal / slow + avoid / stop; EWC anchors the already-safe frames so the new alignment doesn't erode them.
Proportions and figures from the SCAND safe-vs-danger frame analysis (2,460 danger + 2,442 safe frames).