Data quality

… blocking checks passed.

Raw records

…

As downloaded.

Kept

…

Used for every figure on this site.

Set aside

…

Kept plus set aside always equals raw, year by year.

Checks

What has to be true before publishing

Loading the checks…

Set aside

Why records were removed

A record is listed under the first reason it meets. Nothing is dropped silently.

Duplicates

Share of each year's value that was repeated records

Nearly half of 2025's raw value was the same records repeated, some more than 50 times.

Names

Telling one supplier from another

… agency spellings became … agencies, and … supplier spellings became … suppliers. Matching uses fixed rules only, scored against … sampled name pairs (… more were too unclear to judge).

Recall is measured among name pairs that share a first word and look alike. The rules were adjusted once after seeing this sample, so a fresh sample would score somewhat lower.