Agentic systems often rely on a pointwise judge that assigns a single score to a response according to a criterion. Small language models are attractive for this role because they can be served and specialized cheaply, but pointwise labels alone under‑determine the decision boundary: responses to the same query differ along many dimensions, so the true deciding features remain hidden.\n\nThe paper introduces CPA (Counterfactual Pairwise Augmentation). For each pointwise example, CPA synthesizes a counterfactual response intended to receive the opposite label, then frames the accepted pair as a pairwise question asking which response satisfies the criterion. The model is fine‑tuned jointly on pointwise and pairwise data, yet deployed only in pointwise mode.\n\nA second‑order analysis shows that merging separately trained pointwise and pairwise experts moves every parameter direction toward the pairwise expert by the same fraction, whereas joint training can keep the pointwise optimum where the pairwise loss is flat and absorb curvature where it is sharp—precisely the regime CPA is designed to create.\n\nExperiments on six binary criteria using Qwen3‑8B LoRA adapters report a 2.5‑point macro‑F1 gain over pointwise‑only training and outperform model merging at every tested merge weight. CPA also surpasses the strongest zero‑shot frontier judge evaluated, Gemini 3.6 Flash, by 11.8 points.\n\nReview: CPA’s use of same‑source, minimally altered counterfactual pairs enables small models to expose the latent features that drive label decisions, offering a practical way to boost judge performance when annotation signals are sparse.