Thanks for the explanation. I have no doubts about the formula or the calculation steps, but I think there may be an issue with the input data when system is running, this might be edge cases.
In this case, you can see that Edward Y C Tan appears twice in the list (#63 and #88). The #63 entry has both its initial and post-event ratings empty (as shown in the UI). When the system tries to use this player’s rating to calculate other player’s expected score and bonus, it may first pick the #63 record. Since that record defaults to 0 or 100, the result ends up giving a higher expected score but a lower bonus in Jonathan’s case.
I also ran several ratings using the estimator for this section. The large discrepancy only appears when a player faces this re-entered record. In all other cases, differences are minor, usually only 1–3 points.