Different units uses A [1,2,3,4]; Outlier uses A [1,2,3,20]; Constant feature uses A [2,2,2,2]. Base B is [100,150,350,400], multiplied uniformly by Feature B unit multiplier 1..10. Case identities stay P1..P4. Scenarios preserve method and multiplier; prediction changes and Reset restore the current experiment’s baseline. Parameter, scenario and method edits clear stale explanations.

Raw uses output = x. Min–max uses (x − reference minimum)/(reference maximum − minimum). Z-score uses (x − reference mean)/reference population SD, with variance divisor 4. The reference rows are the fitting set here; a future preprocessing pipeline must fit on appropriate training data and reuse its recipe for held-out inputs. No editable unseen inputs or train/test evaluation is implemented.

For zero range/SD, the used denominator is 1 after subtracting the center. Observed constant values map to zero, retaining zero output SD. Otherwise min–max bounds apply to these reference rows; values outside a fitted range need not stay in 0..1. Z-scores can be negative or exceed 1. Neither recipe removes outliers, makes data normal, equalizes distribution shapes or every pairwise contribution, adds information, or guarantees better model performance. “Normalization” here refers only to column min–max, not row unit norms.

All numeric evidence is computed at full precision; displayed values are rounded. Only P1/P4 is compared in the contribution bars, on a common percentage scale. Summing differently measured raw feature numbers is a demonstration of numerical weighting, not a physical units statement. No model training, learned feature importance, robust/quantile transforms, arbitrary point/probe editing or automatic scaling-policy selection is implemented.