An engineer tests a bold hypothesis: whether a swappable, weaker AI model wrapped in a robust software harness can match the performance of frontier-level models. The harness combines deterministic knowledge retrieval, curated tools, verification loops, and structured workflows to compensate for the model's limitations. The approach promises to