Hi @RogerDanilo. There is no good theoretical justification for using 1,000 as a threshold for bad scaling, other than that some (nonlinear) solvers advise to have all non-zero coefficients between 0.001 and 1000 (in absolute sense).

In practice it is often not possible to scale a model in such a way that all non-zero coefficients are in this range and therefore for practical models I often use 100,000 for the bad scaling threshold option.