when we want to find high rank curve over $\mathbb{Q}$ among an elliptic curve over $\mathbb{Q}(t)$ 
,we give some scores for the curves (the curves with bigger points on them over various primes in reduction module p: #$E(F_p)$,i mean $S(E,N)=\sum_{p<N}\frac{2-a_{p}}{|E(F_p)|}*log(p)$,it is briefly called Mestre's sum)
and the curves with bigger scores usually have bigger rank than the original curve over  $\mathbb{Q}(t)$.the specialization theorem help us to be sure that group does't becomes smaller.

 it is interesting that when we do the same thing from  $\mathbb{Q}(t,t')$ to  $\mathbb{Q}(t)$ and try to compute rank of surfaces with conjectural limit ( proposed by Nagao,the special case of Tate's conjecture,proved by Rosen and silverman for rational surfaces) it gives the same rank always, not bigger rank in any cases( i tried an example and compute about 100000 surfaces and compute the limit, all of them was equal)

what's happening?