Hi,
I’m not aware of any paper systematically assessing the accuracy of reaction times in Ibex (or PCIbex) but I wouldn’t be surprised that some published papers reported self-paced-reading experiments designed with Ibex.
What I can tell you though is that you shouldn’t worry about varying speeds of the Internet connections: at least with (PC)Ibex, but I suspect other online softwares do the same thing, all the content of the experiment is downloaded at the beginning (modulo images/audios/etc. in pure Ibex–PCIbex does handle media preloading). So once a participant starts an experiment, they won’t need to communicate with the server until they submit their results at the very end.
The concerns mostly come from each participant’s local environment: do they have lots of tabs open in their browser? Do they have an old, weak processor? What kind of keyboard are they using? If all goes well, most participants won’t encounter performance issues, and then maybe you’ll be able to detect an effect size of 25ms.
If you can afford it, maybe run a small control experiment where you know there to be an effect of comparable size between two conditions to get a sense of what you’ll get?
Hopefully other people here can point you to relevant papers
Jeremy