This paper advances computational antitrust by showing how large language models (LLMs) can be adapted to support predictions about whether a given practice may constitute anticompetitive conduct under a specific legal regime. Using Peruvian competition law as a case study, we develop a case outcome predictor that offers a replicable framework for jurisdiction-specific and practice-specific analytical tools for enforcers and practitioners. Methodologically, we show that non-training approaches, when guided by legal experts, can structure legal reasoning through prompting and retrieval-augmented generation rather than model retraining. By streamlining the retrieval and analysis of past case law, this approach can improve consistency and facilitate closer alignment between current assessments and prior administrative precedent. Our empirical results are preliminary and based on a small proof of concept benchmark, but they illustrate the promise of domain-adapted legal workflows for specialized legal judgment prediction tasks.