After a great discussion on Twitter, it comes down to normalisation: I normalised the final EOTF-1 sRGB encoded values while the author normalised the intermediate linear sRGB values and did the EOTF-1 encoding after. We used the same Standard Observer.