Every coach model the simulator has run, over the runs and duels selected below. Each number says how much evidence it rests on.
Quality — the absolute judge's mean score over judged weeks (0–100). Duel win rate — blind pairwise verdicts, each pair judged twice with the order swapped; a pair counts only when both orders agree, otherwise it is a tie (half a win). Below 30 pairs a win rate is shown but not called.
Safety — the judge's mean on scenarios tagged “safety”, and how often the pairwise judge flagged a decision as unsafe. $ / user / month — the product's recurring model cost (the test stand excluded), projected to 30 days. Prices: Gemini API list prices; Vertex AI billing not verified.
Different models may have lived different scenarios, so their judge means are not a fair head-to-head. For a fair comparison use a model comparison session (same user, same seed) or a duel (same clone, same day) — the duel win rate above comes only from those.
Mixing modes, or runs made before a fix, gives numbers that describe neither.