Summary
After a successful Predict all buildings run (112,689 buildings), the "Predicted N buildings in viewport" counter is non-monotonic: zooming out past roughly z11.2 makes the reported count collapse by ~200x, even though the map is rendering more footprints at that zoom than at any other.
Steps to reproduce
- Run Predict all buildings on a large layer (~110k buildings).
- Zoom out progressively and read the viewport counter at each level, allowing time to settle.
Measurement
Same map centre, 20 s settle at each step to rule out tile latency:
| Zoom |
"Predicted N in viewport" |
Footprints rendered |
| 11.13 |
413 |
109,668 |
| 11.54 |
78,469 |
90,404 |
| 11.80 |
74,988 |
90,404 |
| 12.13 |
59,116 |
93,045 |
| 13.13 |
14,768 |
26,222 |
| 14.13 |
2,903 |
6,308 |
| 15.13 |
933 |
3,072 |
| 16.13 |
322 |
926 |
| 17.13 |
103 |
826 |
From z12 up the trend is sensible. At z11.13 it inverts. The value is also unstable — repeated visits to the same z11.13 camera returned 791, then 413.
Impact
Minor, but the natural first move after a prediction run is to zoom out to see the whole affected area. Doing so reports a few hundred predictions instead of ~110k, which reads as "the prediction run mostly failed". The prediction overlay itself is drawn correctly at that zoom, so the readout directly contradicts what the user is looking at.
Suggested fix
Compute the counter from the prediction store filtered by the camera bounds, rather than from rendered/decimated tile features. Alternatively, suppress the number below the zoom at which it is reliable.
Summary
After a successful Predict all buildings run (112,689 buildings), the "Predicted N buildings in viewport" counter is non-monotonic: zooming out past roughly z11.2 makes the reported count collapse by ~200x, even though the map is rendering more footprints at that zoom than at any other.
Steps to reproduce
Measurement
Same map centre, 20 s settle at each step to rule out tile latency:
From z12 up the trend is sensible. At z11.13 it inverts. The value is also unstable — repeated visits to the same z11.13 camera returned 791, then 413.
Impact
Minor, but the natural first move after a prediction run is to zoom out to see the whole affected area. Doing so reports a few hundred predictions instead of ~110k, which reads as "the prediction run mostly failed". The prediction overlay itself is drawn correctly at that zoom, so the readout directly contradicts what the user is looking at.
Suggested fix
Compute the counter from the prediction store filtered by the camera bounds, rather than from rendered/decimated tile features. Alternatively, suppress the number below the zoom at which it is reliable.