In generative training, models produce an output and penalise its difference from an observed example. With one output per comparison, models capable of producing only one valid image incur the same average penalty as models that have learned the whole distribution. Training objectives cannot discriminate between. Explorative Modeling (XM) produces $K$ outputs per comparison, penalising the closest. They claim exploration is a third pretraining axis. But is it? What does it scale? Here I show the third axis is weakness. Weakness counts how much a model could still narrow what it outputs. How loose your constraints are. Freedom of function, not form. Earlier work proved weakest correct models are likeliest to generalise, beating measures of form like flatness and MDL. I prove the average XM penalty depends on the model only through the chance a single output misses by more than an amount, and exploration raises that chance to the power $K$. For $K>1$, among models producing only correct answers, match probability rises strictly with weakness. For unseen prompts, requirements chosen uniformly at random, the chance of meeting every demand is proportional to weakness. So I ran two experiments. Over 456 runs, raising $K$ increased measured weakness in every primary paired comparison. Hence XM is a means of increasing weakness. I then modified XM, using weakness as a selector vs baseline comparison. Applied to the same candidate pools, weakness raised best-of-eight hit from $0.4131$ to $0.4203$, winning in 19 of 20 worlds over baseline XM, tying the last. Weakness makes XM better. Exploration is a means, weakness the end. The third axis. Always has been.