A separate integrity process tries to prove our data wrong:
- Every build is gated. Records with missing sources, invalid values or duplicates are held back ("under review") instead of shown.
- Weekly sample. We re-fetch a random sample of records straight from the source and compare them field by field. The latest accuracy and its date are on the methodology page. If it misses our 97% target, the page says so.
- Acquisition sweep. Companies that show up in public deal news are removed from feeds.
- Source checks. If a source suddenly returns far fewer records, we stop publishing until a person checks it.
Inferences are only written from sourced facts. Any sentence that names a number we did not see, or guesses about a person, is rejected.