Dear Edwin,

PAML and Hyphy use different models, so some degree of disagreement is expected.
I invite you to read our 2005 paper in MBE (Not so different ...) if you are interested in the details.
That said, a complete disagreement is unusual, so I would caution against placing too much faith in either set of predictions. If your alignment is small or has low divergence, then there may simply be too little power to detect individual sites.

Sergei