Source
Santurkar et al. (2018), LIT-015 — ARXIV-1806.02375.
The same practice as SOTA-004
Same recommendation, same words, different citation: SOTA-004 carries it from LIT-002, the paper that introduced batch normalization and stated the placement; this one carries it from LIT-015, which is about learning rates and does not argue placement at all.
Superseded rather than rejected, because the recommendation is correct and still in force — it is held by SOTA-004, and a reader arriving here should go there. Retiring it as wrong would misrepresent what happened, which was a duplicate import rather than a mistaken belief.
Worth recording how it survived: two practices with near-identical titles, in a registry with no check for that, cited to two different papers so nothing about their sources looked alike either.