Blue is the one colour in this model defined by two absences: a child arrives at it only by inheriting no brown allele and no green one. That is what makes it worth computing, because both gene pairs of both parents have to cooperate for it, while brown is settled at the brown locus alone. Each field contributes one of three numbers — the chance that this pair passes on its recessive allele: 1 for bb or gg, ½ for Bb or Gg, 0 for BB or GG.
The two loci count as independent, so the four chances simply multiply. The defaults are the classic cross, both parents Bb and Gg: ½ × ½ = ¼ for no brown, ¼ again for no green, and ¼ × ¼ = 6.25 % blue. The same structure explains the edges of the table.
A single BB or GG sets its own factor to 0 and rules blue out completely, while two parents who are bb and gg have nothing else to pass on and the model returns 100 %. What the page cannot supply is its own input. Current genetics describes eye colour as polygenic — OCA2 and HERC2 carry most of it, with a dozen further genes contributing — so the tidy gene pair you are asked to choose is a classroom convention, not something readable off an iris or a family tree. Everything downstream of that choice is exact; the choice itself is the simplification.