Published corrections. Superseded figures are retained and linked forward, never deleted — a result page should link to what it replaced. Most research platforms present corrections as though they never happened; here they are a first-class artifact with their own identifier.
No errata published.
Three questions, three answers. One label cannot carry all of them.
Both errata here are numerical corrections, but with different causes — one an execution defect, one a data defect. A future INTERPRETIVE / STATISTICAL case would be a conclusion that was wrong because a statistic was misread, with every figure correct. Collapsing the axes would lose the distinction that decides what to re-check.
Both errata so far were found by the platform's own artifacts rather than by review. Erratum 01 surfaced when execution lag became a declared field and the attribution turned out to credit weights with the return of the day they were decided. Erratum 02 surfaced when the assessment layer began reporting observation counts as a fact, and 3,103 observations over eight and a half years proved impossible.
That is the argument for making assumptions explicit: a hidden choice cannot be checked, and an unchecked choice is how a wrong number survives.