Planverity sits between your application and every model you are allowed to use. It decides what is permitted before it decides what is optimal, and it records both decisions for every request.
Analyse the request, resolve a trusted pool, apply policy, rank what survived, reserve budget, execute, record. Reversing any two of those steps produces a system that looks similar and enforces nothing. The most common inversion is ranking before permitting — it means a high relevance score can reintroduce something your policy excluded.
No model call is made to decide how to route a model call. Token estimation, task classification, complexity scoring and candidate ranking are all local computation. Semantic features exist, are versioned, and are disabled unless they demonstrate repeatable benefit.
Policies, pools and prices are immutable and versioned. You publish version N+1; you never edit N. Editing in place would retroactively rewrite what past executions are recorded as having been permitted, destroying the reproducibility manifests exist to provide.
It does not host your models, and it does not hold your provider balance. Planverity is bring-your-own-key: your credentials are resolved at the execution boundary and the provider bills you directly. Managed access and a credit ledger are designed but not built.