AI models and connection
Distinguish a chosen provider, a verified connection, and an actual model run.
On this page
What the connection light means
The Company composer checks the existing main-intelligence metadata endpoint on load for the verified current account. Neon green means that check confirms an available model connection. Red means it is not confirmed or available for this participant; a pending check is shown separately.
That metadata check sends no Company or practice draft and performs no inference. It does not prove that a particular answer is correct, that a worker is running, or that an outward action has happened. Signed-out readers do not borrow an owner's connection.
Selection is different from execution
The shared picker retains its existing device model preference. Selecting a model does not call it. A separately authorized processing path determines which source, provider, operation, limits, and receipt apply.
Model availability and defaults depend on the server's current configuration. This guide does not assert that a provider is configured, that a paid run was made, or that a staged implementation has been released. Inspect the relevant interface and its current response.
Jev and typed judgments
TypeSafe describes Jev as its typed System One model. Its official interface asks bounded Choice, Score, or Noul questions against supplied state and returns structured judgments. Provider confidence describes that judgment under its method; it is not independent validation, consent, or authority.
Yawn's existing Company purpose and move review uses the Forman Brain typed request and receipt contract. Its verified owner explicitly approves the exact current context with pinned jev-1.13.0. A saved review shares the existing orientation model-run ledger. A draft review is not persisted; its five-minute process-local replay cache is not a durable or distributed allowance. No returned judgment chooses or executes a Company move.
The Confidence panel at /typesafe-ai/jev offers a separate, one-question review over deliberately selected saved text. It uses the same server-only TypeSafe credential and established model-run ledger, while retaining its own exact rubric, model, usage, source hashes, revision and explicit one-call approval. This panel also requires the existing capture-execution flag. A stale result stays attached to its original basis; retrieval never starts another call. Opening the owner panel may check provider model metadata, which is separate from inference.
The provider documents experimental images and Box with per-request opt-in. The current Company and Confidence consumers send text only; Forman Brain's image manifest validation does not decode or upload images. Provider capability, connected application capability and an actual retained run remain distinct.
Versions and limits matter
Provider aliases can point to different revisions. At the official reference check on October 16, 2026, TypeSafe listed jev-1.14.0 as the newer explicit revision while jev-latest still pointed to jev-1.13.0. Inspect the provider's current model reference rather than assuming an alias means the newest revision.
Text-state input, accepted questions, availability, and limits are provider-specific. Keep the model and revision with a result, retain uncertainty, and do not promote a typed judgment to a canonical Company action.