What I measured when I put Jev next to gpt-5-mini

I ran the same support tickets and agent tool calls through Jev and gpt-5-mini. On one ambiguous ticket, asked twenty times, gpt-5-mini answered billing fifteen times and technical five. Jev returned both probabilities in one call.