Agent evals have a hidden problem: The judge itself can be unreliable. A new experiment tested Jev against GPT-5.6 Luna, Terra and Claude Sonnet 4.6. The results: → 500/500 binary decisions matched the human oracle → 92–913× lower score variance → 0.44s average latency → $0.00035 per evaluation Jev doesn't generate a paragraph and then turn it into a score. It makes a structured decision directly. That could change agent evaluation economics. Cheaper, faster judges mean teams can evaluate more traces, run more regression checks, and tighten the feedback loop around every agent update. The important caveat: this was a narrow early experiment. But the direction is fascinating: Reliable agents may require better judges, not just better models.