A third of public ownership data names the previous owner

We were trying to close a gap. 29,000 assets in our dataset carried no owner, and two public sources looked like they could fill it: the UK Renewable Energy Planning Database, published quarterly by DESNZ under an open licence with an "Operator (or Applicant)" column, and Wikidata, which records an owner or operator for 4,462 European power stations.

Both matched well. REPD matched 85% of unowned UK assets once its OSGB36 coordinates were converted to WGS84. Wikidata matched on name, capacity and proximity across the continent. Either would have added thousands of owners overnight.

Before merging, we ran the same matcher against assets whose owner we already knew and checked how often the source agreed. That is the entire method, and it is the only reason neither dataset is in this product.

Both disagree on about a third

A third is not noise. Random error would scatter across plant types and sizes. These do not scatter — they cluster on assets that have changed hands.

The failure mode is the same in both

REPD records the operator or applicant at the point of planning. Wikidata records what a volunteer entered, usually when the plant was commissioned and newsworthy. Neither is maintained transaction by transaction, because neither was built to be: REPD tracks projects through the planning system, and Wikidata is an encyclopaedia. Both do their own job well.

But it makes them accurate for exactly the assets nobody is asking about, and stale for exactly the ones people are. A plant that has changed hands is a plant somebody paid for, diligenced, and now wants comparables on.

What we did instead

Neither source was merged. A wrong owner carrying a citation and a confidence score is worse than a visible gap: the gap is obvious and the fabrication is not, and the people who would catch it are the people who did the deal.

The candidates became a research queue instead — largest assets first, with what each source claims and whether two sources agree. Where two independent sources do agree, the staleness that breaks each one alone would have had to happen identically in both, which is a far stronger signal than either on its own.

The assets closed since were done one at a time and cited to the owner’s own announcement: Kruonis to Ignitis Gamyba, Ardnacrusha to ESB, Cowley Complex to Sonnedix. Slower, and it is the only method that survived its own test.

How to check an ownership source in an afternoon

Our Wikidata matcher ships with a validate mode that runs this held-out test and prints a verdict, so the check happens before the merge rather than after. The result that stopped us took about an hour to produce. The alternative was 1,890 confident, sourced, systematically wrong records.

• –Take records where you already know the answer — not the ones you are trying to fill.
• –Run the same matcher you would use in production. One tuned for the test proves nothing.
• –Compare names forgivingly: strip legal forms, accept initialisms. Err towards agreement.
• –Read the disagreements individually. A pattern in them tells you what the source really records.
• –Only then decide whether to merge, and at what confidence.