Usual is the median failure rate on the same weekday over the previous eight weeks. Score is the gap divided by a robust spread (1.4826 × the median absolute deviation), so ±4 is roughly four standard deviations and is unlikely to be noise.

Pipeline that day comes from the data health checks. If it is not OK, the move may be a data problem, not a change in behaviour, and should be checked before anyone explains it. That question is what the twinfault benchmark tests.

The largest spikes are explained, with evidence, on the Investigations page.

Failed transactions emit no transfer events, so failures are only visible in the transactions table, from calls sent straight to the token contracts. Days with fewer than 500 direct calls for a token are left out.