The cap is reportedly at ± 10 net wins (see below) – which should translate to ±20k, per your observation.
However, back in March, Kerridoc reported that mhalttu had reported that the adjustment is multiplicative. His illustrative “not the actual numbers” was a ±2% of the (unmodified) alliance score per net win.
… and of course, that may have chanced since March …
I believe I saw them mentioned on Wikipedia as examples of Elo rankings.