A prettier map still lit up the wrong district
After shipping more detailed Korean boundary tiles in Wimemo, Seoul + Busan photos could still light the wrong district.
That was the useful embarrassment: prettier map, worse truth model.
The bug wasn't boundary detail anymore. It was that I was letting render geometry and visit evidence share the same source.
- Korea dataset: 17 ADM1, 228 ADM2, 876 vector tiles (~696 KB)
- the Atlas looked better after the tile pipeline
- but a temporary card centroid could still imply a place the user never visited
The fix was boring but durable: only source-photo coordinates can prove a visited area. Card centers, cluster centroids, and visible labels can summarize evidence, but they can't replace it. We now persist normalized region keys like KOR|KR-11 and keep render tiles for drawing only.
Reusable lesson: a better UI artifact is not automatically a better source of truth. If a piece of data exists to help the screen render, assume it will eventually lie to your product logic.
What UI or visualization artifact in your product looked "good enough" until it started making behavioral decisions it was never meant to make?
