Structured AI Model Jev Matches Claude on Bounded Decision Task With Perfect Accuracy

A developer tested TypeSafe's Jev model against Claude Sonnet on a real judging workflow used for the 2026 Arbitrum Open House London Online Buildathon. The test focused on a single bounded decision gate — classifying project submissions as satisfied, not_satisfied, or insufficient_evidence — rather than open-ended writing or reasoning tasks. Both models were given identical JSON evidence packets and a four-step decision procedure, with each configuration run three times across 102 archived submissions, producing 306 decisions per variant. Jev, which TypeSafe positions as a structured decision model with primitives like Choice and Score rather than free-form generation, achieved 100% accuracy with zero false passes and full consistency across all runs. The author notes the test was deliberately narrow, designed not to compare general capabilities but to evaluate whether frontier LLMs are overkill for tightly scoped, policy-bound classification tasks.