One overloaded category (every "door" = an "exit") leaks the bug into every consumer; cross-reference existing metadata to triage without re-capture
A building-training app let users mark door/window sweeps as "exits" via one panel. But users (correctly) marked EVERY door, including interior ones (bedroom/bathroom/closet), and the code treated all of them as building egress. The lesson in modeling: when one tag conflates two real-world concepts (interior door vs exterior egress), the conflation leaks into EVERY consumer of that tag -- here three independent ones: the route planners usable_exits (interior doors became valid route endpoints), the APIs exitSweepIds list, and the clients drill-completion trigger (walking to a bedroom door ended the drill early). Fixing one is not enough; grep every reader of the field and apply the same predicate (we gated all three on type in {door,window}, added a distinct interior-door type that is recorded but never egress, and relabeled the capture UI buttons so intent is unambiguous: "Exit door" / "Interior door (not an exit)"). Second, reusable technique: when a user has already done manual labeling that now needs re-classification, DONT make them redo it -- cross-reference the marked items against OTHER metadata you already have to triage automatically. Here, joining each exit sweep to its room label instantly flagged the likely-interior ones (door whose sweep sits in a room labeled Bedroom/Bathroom/Landing = almost certainly interior; door in an unlabeled entry area = probably the real exterior exit). Present that table and ask for one confirmation instead of a full re-capture. Keep the final judgment with the user (which door is the front door is their domain knowledge), but do the heavy filtering for them.