A non-automatic (!) application of Gosper's algorithm evaluates a determinant from tiling enumeration (Q1415038)

The authors prove the determinant evaluation \[ \begin{multlined} \det_{1\leq i<j\leq n}\Biggl( \binom{x+y+j}{x-i+2j}- \binom{x+y+j}{x+i+2j}\Biggr) \\ =\prod_{j=1}^n \frac{(j-1)!(x+y+2j)!(x-y+2j+1)_j(x+2y+3j+1)_{n-j}} {(x+n+2j)!(y+n-j)!}, \end{multlined} \] where \((a)_n=a(a+1)\cdots(a+n-1)\). This determinant, which was previously proved by the same authors using a combinatorial argument combined with a known determinant evaluation, counts the number of lozenge tilings of a hexagon with cut-off corners. Here the shape of the cut-off hexagon is parametrized by \(x\), \(y\) and \(n\). The method of proof is that of ``identification of factors'', which -- thanks to the work of the second author -- has become one of the standard techniques for evaluation determinants. Not at all standard, however, is that in the course of their proof the authors need to establish the following transformation. Let \(P_l(x,y)\) be the polynomial \[ P_l(x,y)=\sum_{r=0}^{2l+1} a_r (x)_r(-y)_{2l+1-r} \] with \(a_r\) the coefficient of \(z^r\) in \((z^2+z+1)^{l-1}(2z+1)(z+2)(z-1)\). Then the transformation states that \[ \begin{aligned} &\hphantom{=}\sum_{k=0}^{i+l}\frac{(k+i+l+1)_{i+l-k}}{(i+l-k)!}P_l(2i,i+l-k) \\ &\qquad \quad \times (-2y-3i-l+2j-k+1)_{n+k}(y+k-j+1)_{n-k} \\ &=\sum_{k=0}^{i+l}\frac{(k+i+l+1)_{i+l-k}}{(i+l-k)!}P_l(2i,i+l-k) \\ &\qquad \quad\times (-2y-3i-l+2j+k+1)_{n-k}(y-k-j+1)_{n+k} \end{aligned} \] for \(1\leq l\leq n-i\). As it turns out, both sums appear to be Gosper summable. For example, for \(i=3\), \(j=2\), \(l=1\) and \(n=4\) they sum to \[ -13440(y-2)(y-1)y(y+1)(y+2)(y+3)(2y+3)(2y+5). \] However, the authors have been unable to find an explicit expression for the sums for arbitrary \(i,j,l\) and \(n\). Despite this difficulty, they show that using Gosper's algorithm -- which is designed for automated indefinite summation of hypergeometric sums -- in a non-automatic fashion one can prove the above transformation without ever explicitly finding the actual evaluations. Hence the non-automatic (!) in the title of the paper.