. (2026). “From Slogans to Classifiers: Advertising-Inspired SDG Boundary Engineering”. 11th International Conference on Web Intelligence, Machine Intelligence and Semantics (WIMS 2026), accepted: 2 July 2026
Classifying documents against the United Nations Sustainable Development Goals~(SDGs) is essential for tracking progress toward global sustainability targets, and hard because the goals overlap by design. We present an advertising-inspired framework for decision boundary engineering that turns model shortcuts and prototype cues into synthetic training signal through four mechanisms, catchphrase anchors, contrastive antidotes, poetry anchors and contrast sets. Each is evaluated with ModernBERT-large on an expert-annotated and a community-annotated corpus, against a continued-training control that isolates a mechanism's effect from the effect of training longer. We report two findings. First, which mechanism helps depends on the corpus: antidotes give the only significant gain on the community-annotated corpus (+0.65 percentage points (pp), 95\% CI [+0.10, +1.21]), whereas on the expert corpus the antidote--catchphrase combination (+1.24 pp) and Manifold Mixup (+1.09 pp) lead; the gains are small, on the order of run-to-run variance, and none wins on both corpora. Second, measuring class separability numerically rather than reading it off a projection, we find it does not predict macro F1 (r=+0.67, p=0.15 on the expert corpus; r=-0.31, p=0.55 on the community corpus), and that the corpus where antidotes help most is the one where they reduce separability. We report this dissociation in place of the geometric explanation we set out to give.