Refusal comes first
Region, provider, data classification, cost ceiling and tool permissions are evaluated before any ranking. A model that scores perfectly on relevance cannot be reintroduced by that score if policy excluded it.
Planverity routes models and tools under your policies — optimising cost, quality and reliability while recording exactly what ran, why it was chosen, and what it cost.
Your AI stack has gateways. Planverity gives it judgment.
const client = new OpenAI({
apiKey: process.env.OPENAI_API_KEY,
});const client = new OpenAI({
apiKey: process.env.PLANVERITY_API_KEY,
baseURL: "https://api.planverity.ai/v1",
});Your request and response handling do not change. An existing client does not know the difference.
The difference between a gateway and a control plane is what happens before the optimisation, not after it.
Region, provider, data classification, cost ceiling and tool permissions are evaluated before any ranking. A model that scores perfectly on relevance cannot be reintroduced by that score if policy excluded it.
Spend caps use reservations, not counters. Two requests reading the same remaining budget cannot both proceed and collectively overspend it — the check and the reservation share one transaction behind a lock.
The manifest records the plan, the policy and pool versions in force, the candidates refused and why, every attempt including fallbacks, and what was billed — written before the first provider call.
Available → policy allowed → relevant → selected. Never relevant → authorised. On a 60-tool catalogue this took forbidden-tool exposure from 798 to zero.
These come from our own benchmark runs. The uncomfortable results are here too, because a routing product that publishes only its wins is not offering evidence.
Planverity is Pareto-efficient but does not dominate a good mid-tier model. Against always-Sonnet it 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.
Semantic routing is built and shipped off, after failing to beat deterministic ranking in three independent measurements.
The console distinguishes measured from estimated everywhere, because the difference is the product.
Verified quality needs a grader on live traffic. Until that exists the console says so, rather than rendering an unmeasured score as 0 or as a failure.
When no alternative model was feasible, savings show “no alternative existed” — never $0, which would claim the alternative cost nothing.
An endpoint with no price on file fails closed under a cost ceiling. Cost estimation returns null, never zero, so unknown cannot masquerade as cheap.
The demo organisation includes an execution where two stronger models were refused — each citing a measured p90 output length against the caller’s token budget.