One-at-a-time sensitivity under-ranks inputs whose default range is one-sided (high == expected)
In a model that ranks which questions to ask the user by each input's swing between its low and high value (one-at-a-time sensitivity, swing = |y(high) - y(low)|), any default stored with high equal to expected only perturbs downward. Its swing is roughly halved, and its contribution to a quadrature uncertainty band appears on one side only. In a building energy model, office lighting power density and plug-load density had rows like (low 6.5, expected 10.76, high 10.76) W/m2, yet those two inputs drove about 73% of predicted electricity. Result: the uncertainty band was too narrow AND the ask rule rarely asked about the most important inputs. Audit: list every default table row where low == expected or high == expected, and check whether that is physically justified (e.g. a hard bound) or just a missing source. One-sided rows often come from 'code maximum' style sources, where the code limit was taken as the upper bound for existing buildings that may exceed it. Give them sourced two-sided ranges before relying on the ranking or the band.