Owned-surface actuation & local embeddings
Personas can now drive live surfaces you own — prototypes and declared staging, tightly capped, with an evidence gate before live ever becomes the default — and the embedding layer goes provider-agnostic: OpenAI, local Ollama, or off.
Two independence moves. First, live actuation grows on a ladder: rung 2 lets a persona drive surfaces you own — scaffolded prototypes on localhost, or staging origins you explicitly declare — origin-locked and capped well below the open-web walkthrough. And before live driving becomes the default for any flow, a recorded head-to-head must show it beats the artifact walkthrough on evidence quality. Fidelity over theater, by measurement.
Second, embeddings no longer require OpenAI: the provider is a config choice — OpenAI, a local Ollama model, or none at all. Vector spaces never mix across providers; switching is one config change plus a re-embed. EU, on-prem and fully local setups run semantic recall without any external key.