The eligible sets are in the same compressed sparse row form that presents a graph to a machine elsewhere in the archive, for the same reason: it is the adjacency-array format an algorithm would actually be handed, with nothing precomputed. The blocks are not required to be sorted and repetitions are not forbidden. Leaving those conditions out admits more words and therefore strengthens, rather than weakens, every claim about programs reading the format.

Here magnitudes stop being free. The processing times, deadlines and weights are entries of the word, so a claim about a program reading it has to say that they are words — which the fitting conditions of $ParameterizedComplexity$ do, once, as an explicit inequality against $2 ^ w$. A size measure that counted only the number of jobs and machines would make the weights of a reduction's output invisible, and a running time stated against it would not be a claim about anything a machine does.

Cells are read with $List.getD$, which returns $0$ outside the word; the length condition pins the word down completely, so the default is never reached at a position the other conditions constrain. Unlike a graph encoding, the length is not determined by the header alone — the target array is as long as the last offset says — so $length_eq$ reads that offset rather than a declared edge count.

The threshold is appended last, so that the instance block sits at the same offsets whether or not a threshold follows it, and the split of the word into the two parts is determined by the word rather than chosen.