Vicnaum deserves a medal for clarifying stuff ! :- )
Yes, some PBR models (like the one in Unreal engine 4 derived from Disney's paper) use inverted specular reflectivity values because they wanted UNIFIED model for specular reflectivity across both metals and not metals.
Why is this logical: (I am not advocating we need this actually at all, but just writting argument against "all the mainstream renderers do something, so it's obviously right" fallacy type of argument above).
1)Metals don't translate well using only the simplified fresnel formula (only n-number), so they end up looking "incorrect" (of course, depends to what point at which case), and to avoid using the complex model (n,K), which
for example Maxwell has, they ditched it altogether and just let you specify base reflectivity and metal-ness (0 or 1, whether the material is metal, or not metal). The rest is adjusted internally then.
This also lets you lineary scale the specular value from non-metal (which will be at 0.02-0.04) to something like 0.6 for chrome or 0.9 for aluminium. One value for all material types, instead of two values.
2) If we discount refractiveindex.com , it's easier to get measured values in inverted fashion. It's still potato values, but easier to use, because again, we are only using ONE value (fresnel is computed internally aftewards).
From 0.02-0.04 for non-metal likes plastics, paints,etc.. to 0.2-0.95 etc.. for metals. This is both easier to custom modelling the curve from ref index web, as well as inputting 2 values for full fresnel n/K (too scientific).
Both points above become strong positive when it comes to authoring textures. The whole workflow was intended so you can paint your textures more universally. You only paint single specular map which will only apply with metals,
and will contain both color and intensity (in the scale above,0.2+) but will be ignored for non-metals, which will return to default 0.04 (1.52 IOR as Vicnaum wrote). Both will keep 1.0 at grazing angle and their variance will be driven
by single another map, roughness. It's very simple model actually. I suggest everyone to play with Unreal to get the feel of how it works. It's not more abstract, it's more user-friendly and logical.
It doesn't make creating material "like hell", nor does it look like anything Modo does. It's the same stuff. It's actually easier, and it prevents people from creating "incorrect" materials (metal with too high albedo because you use diffuse, but incorrectly low specularity because you used random IOR number like 8-12 instead of unique curve which varries a lot for metals, and can only be simulated by the respective curve, or full fresnel formula, n/K (like Maxwell).
All the other stuff applies same way, there is no "capping" of specular reflectivity, it always go to 100perc. but the curve can be so steep that coupled with rough surface, the material simply appear mat (and shader lambertian)
Last but not least to counter the "People have no-problem creating photo-realistic materials":
Well, I would say they do.. 99perc. of people including me have that problem. But the problem is not so much that's not possible now. It is just fine. But that it could be refined to allow "easier" creating of them, by a model that very much
navigates you to correct result and avoids the possibility of having physically incorrect materials.
Which the current model promotes, the fact the you can cap reflectivity for non-metals means that vast majority of people have materials that are less reflective then they should. Instead, they often compensate by too high diffuse values or generally upping exposure for whole image and then complain on forums of "flat, washed out" look.
