Clapham Digital · Case study

A public data pipeline that checked its own work

Nowhere in Particular is a driving game that points you at a place and tells you nothing else. Under it sits a corpus of 682,560 markers, merged from three public sources and published under an open licence.

What it is

You get pointed at a road you’ve never turned down. No route, no map line, no picture. You work out the roads yourself. It’s live on the App Store and Google Play, and the app page has the details.

The game is only as good as its map of places. That map merges three public sources: NOAA National Geodetic Survey survey marks (613,497 records), National Register of Historic Places listings from the National Park Service (67,141) and OpenStreetMap (1,922). The whole file is published under the Open Database License, all thirteen fields documented.

The hard part

My own check found most of my corrections made things worse.I checked the coordinate corrections against an independent source, Wikidata. Of the 2,788 it could judge, 1,629 (58.4%) had moved the pin further from the truth than the National Park Service point they replaced. The step had been reporting an improvement rate while it made the data worse.
So the rule changed.Any one independent source can refuse a correction. It takes two to move a point: the Census geocoder and a second witness have to agree, and nothing moves more than 1,999 metres. Only 1,045 corrections are published now. The second witness is Wikidata for 961 of them and Nominatim for 84. Nominatim is a geocoder too, so for those 84 the independence is weaker, and the data page says so.
Licence-clean on purpose.Address matching runs through the US Census geocoder, not a commercial one. Its output is a US government work, so it’s in the public domain. A proprietary geocoder’s terms would have made the corrected coordinates impossible to publish.
One source of truth.The download is exported from the same bundle the app ships. There’s no version of the corpus the app has and you don’t.
It says what it doesn’t know.How reachable a marker looks is known for 18.44% of records. The page spells out that a missing value means nobody has classified it, never that it’s safe. It also lists which of its own numbers it can’t fully reproduce.

What it means for your project

Most business data work has this shape. Pull from messy sources, merge them, decide how far to trust the result. I check the output against something independent and write down what’s allowed to change. Getting a spreadsheet, a CRM and a website to agree is the same job.

Have something that needs building, fixing or checking? Email [email protected] or call or text (614) 835-7163.

All case studies → Blog →