We study non-Boolean PCPs that have perfect completeness and read
three positions from the proof. For the case when the proof consists
of values from a domain of size d for some integer constant d
>= 2, we construct a non-adaptive PCP with perfect completeness
and soundness d^{-1} + d^{-2} + epsilon, for any constant
epsilon > 0, and an adaptive PCP with perfect completeness and
soundness d^{-1} + epsilon, for any constant epsilon > 0.
These results match the best known constructions for the case d =
2 and our proofs also show that the particular predicates we use in
our PCPs are non-approximable beyond the random assignment threshold.