All of your methods are finite-element approximations of what could be considered a truth simulation. So, an inter-comparison between them, as you are doing, will let you know that the results are deviating from each other. However, they wont tell you which of them is more precise as there is no benchmark.
EnergyPlus, if I not mistaken, traditionally uses radiosity-based split-flux method based on DELight.
Daysim uses the daylight-coefficient method based on Christoph Reinhart’s research from 1999/2000.
Honeybee[+] tweaks Reinharts workflow slightly by considering more accurate sun positions.
Have you tried doing a point-in-time radiation study with Radiance on your model (with gendaylit and not gencumulativeSky)? Within software models, thats probably the closest you’ll get to empirical truth. There have been few others on this forum who’ve done that comparision(although I think they were comparing illuminance and not irradiance).
With regards to the accuracy of EnergyPlus, LBNL recently did a comparative validation using multi-phase methods.