How accurate the API is
Each figure is measured against a named truth set and says how many cases it rests on and when it was measured. Figures that miss their own band stay on the page; findings describe a known limit without a verdict. Edition 1, 3 October 2026.
Place search
- The right place among the first three answers 74 %: 148 of 200 200 cases, measured 3 October 2026, against 200 place and outdoor-name queries in several languages and scripts (Wikidata coordinates (CC0), hand-picked, v1). A pass is a top-3 answer within a radius set by the kind of place (1 km for a peak or hut, up to 25 km for a trail). No band declared yet; the classes show where search is weak (other-language names, exonyms, typos).
By class
Figure Cases exact 89.5 % 57 multilingual 38.5 % 26 diacritic 71.4 % 28 punctuation 70.6 % 17 exonym 23.1 % 13 ambiguous 93.8 % 16 typo 53.8 % 13 golden 100 % 30 - The right place as the first answer 64.5 %: 129 of 200 200 cases, measured 3 October 2026, against 200 place and outdoor-name queries in several languages and scripts (Wikidata coordinates (CC0), hand-picked, v1).
By class
Figure Cases exact 75.4 % 57 multilingual 26.9 % 26 diacritic 60.7 % 28 punctuation 64.7 % 17 exonym 15.4 % 13 ambiguous 75 % 16 typo 53.8 % 13 golden 100 % 30
Answer time at the server on 2026-10-02: 95 of 100 calls under 100 ms (budget 500 ms), measured in one region, Helsinki; add your own network time. Live figures: status page.
Analyse a route you already have
- Length measured on the line as sent 100 %: 30 of 30 30 cases, measured 3 October 2026, against 15 race courses and 15 trail stages with official length and climb (organisers' and operators' published figures held with provenance by TrailSplits, v1).Band: within 0.5 % of the line's own length (met).
- Climb on hilly lines (10 m per km or more) 56 %: 14 of 25 25 cases, measured 3 October 2026, against 15 race courses and 15 trail stages with official length and climb (organisers' and operators' published figures held with provenance by TrailSplits, v1).Band: each line within 10 % or 50 m of the official climb. Median ratio to the official climb 0.95. Official race climbs are often barometric or rounded.
- Finding: Extra climb read on flat lines 2.77 m 5 cases, measured 3 October 2026, against 15 race courses and 15 trail stages with official length and climb (organisers' and operators' published figures held with provenance by TrailSplits, v1). Metres of climb per km above the official figure on flat city and coastal courses (median). The terrain model reads bridges and waterside edges as climb.
- Typical difference from the published walking time 15.4 %· 8 countries checked against operator pages 221 cases, measured 3 October 2026, against 221 hiking routes with walking times published by their operators (OpenStreetMap route tags (ODbL), each country checked by hand against the operators' own pages, osm v1). © OpenStreetMap contributors.Band: median 15 % or less (not met). Published times are local formula outputs; in Switzerland, where the model was calibrated, the difference is smallest.
By country
Figure Cases Switzerland 6.2 % 40 Germany 12.4 % 31 France 15.1 % 40 Britain 25.2 % 9 Greece 27.4 % 40 Croatia 16 % 40 Hungary 9 % 13 Sri Lanka 8.9 % 8 - Difference for 4 of 5 routes 27.5 %· 8 countries checked against operator pages 221 cases, measured 3 October 2026, against 221 hiking routes with walking times published by their operators (OpenStreetMap route tags (ODbL), each country checked by hand against the operators' own pages, osm v1). © OpenStreetMap contributors.Band: 80th percentile 25 % or less (not met).
- Finding: Our time against the published time, typical direction −11.9 %· 8 countries checked against operator pages 221 cases, measured 3 October 2026, against 221 hiking routes with walking times published by their operators (OpenStreetMap route tags (ODbL), each country checked by hand against the operators' own pages, osm v1). © OpenStreetMap contributors. Negative: our times are shorter than the published ones.
Answer time at the server on 2026-10-02: 95 of 100 calls under 2000 ms, measured in one region, Helsinki; add your own network time. Live figures: status page.
Calls for this edition came from Server 1 (OVH, Canada), internal, labelled ws11; weekly run 05:49–06:03 UTC. Figures from truth sets that share a source with what they judge are left out, and so are truth sets still being sampled.