The positional error a remapped municipality carries, by what attests it
Source:R/geocode.R
nar_remap_uncertainty_m.RdWhen [normalize_address()] hands back a municipality that is not the one the string named, the place that was searched was chosen by the gazetteer rather than asserted by the input. `mun_remapped` says that happened; `mun_evidence` says on what grounds, and the two populations it separates are far enough apart that one floor cannot serve both.
Measured against Nova Scotia's PVSC assessment points – the one reference established to be independent of NAR – over the 32,512 exact unambiguous building matches in a 40,000-address sample:
| `mun_evidence` | n | p50 | p90 | >5 km | floor | | — | —: | —: | —: | —: | —: | | `kept` | 30,045 | 10.2 | 56.9 | 0.05 | `copostal` | 1,632 | 7.5 | 50.6 | 0.80 | `csd` | 91 | 14.5 | 87.2 | 0.00 | `unattested` | 560 | 12.9 | 83.3 | 1.79 | `untestable` | 184 | 32.0 | 327.5 | 1.63
**The attested classes get no floor, because there is no spread to report.** Pooled, `copostal` and `csd` sit at p90 52.6 m over 1,723 rows – *below* the 56.9 m of the rows whose municipality survived the parse. A remap the register itself vouches for, whether by two names sharing a postal code or by one name being the other's census subdivision, is positionally indistinguishable from no remap at all. Reporting metres against it would be inventing them.
**The unattested and untestable classes pool at p90 118.2 m over 744 rows**, and that is the constant. `untestable` is the wider of the two and is *not* fined by [nar_gazetteer_sql()] – refusing it costs 119 exact matches for 2 errors past 5 km – so the floor is the whole of what is done about it.
`inferred` – the string named no municipality and the gazetteer supplied one – takes the same floor **unmeasured**. PVSC always carries a city, so the Nova Scotia corpus contains none of these rows and cannot price them. Grouping it with the unattested is the conservative reading, not a measured one.
Two things about all of these matter more than their size.
They are *disagreements*, not error budgets: each contains NAR's own distance from PVSC, which the `kept` row shows to be almost all of it.
And none of them describes the tail, which is where the remap risk actually lives. An unattested remap lands more than 5 km out 1.79 untestable one 1.63 at a distance no 90th percentile of any of these populations reports. A caller who cannot tolerate a kilometre-scale error should filter on `mun_evidence` itself. These floors exist so that `uncertainty_m` stops reporting 0 m on the populations where 0 is least true, not so that they can be read as bounds.
See `inst/notes/nova-scotia-pvsc.md`.