The ELO rating is based on the assumption that the expected score of a player A against an opponent B is:
E_A = 1/ (1 + 10^((R_B - R_A)/400))
where R_B and R_A are the ratings of B and A, respectively.
Now, what could be done is that based on the data you give a rating average is computed and the score (per game) that was obtained is computed, this is plugged in for E_A and R_B, respectively. Then, one solves for R_A. [Given that only the global score is entered, whatever is done should be based on the average.]
In other words the hypothetical rating is computed that would yield an expected score matching the actual score obtained. This is how I would interpret the description on the site.
This does however not match what the site does or there are numerical errors.
Maybe they use a different formula for the expected score.
To get a definite answer you might need to contact the maintainer of the site.