Mutable prediction tables undermine reproducibility, a football analytics team finds

A football analytics platform measured its Postgres predictions table and found that reproducibility depends on whether stored rows match what existed when a call was made, not just on model seeds. Of 46,192 rows, 46,181 were modified more than a minute after creation, and 35,013 settled rows were changed after their outcomes were graded. The team has no audit table holding prior versions, so it cannot tell whether those later edits were routine metadata updates or changes to the predictions themselves. The author argues this is a schema problem, since the table allows updates to any unprotected column.