Vendor tiers are a claim. Planverity treats them as a prior with deliberately wide uncertainty, then lets graded trials move the estimate — including against the vendor's own claim.
Routing compares a lower-confidence bound on quality. Wide uncertainty pushes that bound down, so an unproven route cannot satisfy a high floor on optimism alone — the router escalates or refuses instead of routing cheap on unearned confidence.
Measured outcomes update a Beta-Binomial posterior seeded by the tier prior. A hard 'use the raw success rate once n is large enough' cutoff makes the estimate lurch, and twenty successes in a row would otherwise assert perfection. Evidence must be able to move the estimate against the tier, or the tier is an unfalsifiable claim.
Mock runs are excluded — a mock echoes the prompt, so its grades describe the harness. Infrastructure errors are excluded — a rate limit is not a quality failure, and counting it as one makes an unavailable model look incompetent. And no model grades another: these labels calibrate the quality model, so grading with a model would make it circular.
Against always-Sonnet, Planverity offers a dial — 95.1% success at half the cost, or equal quality at about 5% more — not a free lunch. It beats always-premium and a hand-rolled task-tier heuristic, and that heuristic is itself dominated by simply using one balanced model. When one model is near Pareto-optimal alone, routing has little room to add value.