# Capability & launch matrix

Separate the requested product scope from verified service behavior.

## Observed on the existing backend

Authenticated model discovery, a tiny chat completion, chat SSE and a tiny Responses request have passed. These are bounded probes, not broad benchmark, tool-use, load or availability proofs.

## Account and history behavior

The platform supports owned keys and durable admission quotas. Signed-in users can explicitly save, reopen, update and delete chats; unsaved chats remain tab-only. History is isolated by web account, not an API key. Saving includes messages and reasoning and requires consent; there is no automatic persistence.

## Still required for the full product

Public inference, verified email onboarding/recovery, reconciled token usage and billing, full protocol parity, production abuse defenses, external load and rollback remain launch gates.

## Extended API scope

Messages and Responses have bounded local native-GPU protocol/tool-loop evidence and official-client tests; public deployment and broader certification remain open. Text continuation is separate from FIM: suffix is rejected and fim_supported is false until correct middle insertion is proven. Genuine FIM, files, search/fetch and user balance still need their own implementation and exact contract tests. Requested parity is not replaced by documentation or a route that merely returns success.