Every demo run of my eval harness scores 82/100. That's the point.

A demo score that moved around would tell me nothing. How I built an agent-eval harness around three scoring dimensions, 14 deterministic output checks, and explicit limits.