Triaxial evaluation framework (Pass@1, CxSelf, CxJudge) across four architecture conditions. Automated pipeline with model-specific routing for specialist, generalist, decomposer, and judge roles.
FindingGeneralist leads execution (Pass@1: 0.911 vs 0.862); Specialist leads complexity awareness. Naive Hybrid regresses 0.133; guarded variant recovers 87%. A trade-off invisible to single-axis evaluation.