Over-Engineered Parity Classifier Reveals How Data Representation Shapes ML Results

A developer revisited a deliberately complex machine learning classifier built to solve a trivially simple problem — determining whether an integer is odd or even — and discovered significant flaws in the original experiment's design. The initial version reported 69.67% accuracy but suffered from label leakage and a misleading claim of being fully unsupervised, since cluster-to-parity mapping still relied on labels. After rebuilding the experiment with a stricter evaluation protocol using integers from 0 to 10,000 and proper train-validation-test splits, the revised model achieved 84.26% held-out test accuracy. Crucially, masking the least significant bit — the single bit that actually encodes parity — caused accuracy to drop to near-chance levels at 48.15%, confirming the model was exploiting existing binary structure rather than learning an abstract arithmetic rule. The project ultimately demonstrated that data representation choices, such as left versus right zero-padding, can dramatically alter what a model is able to learn, independent of the underlying problem.