A chat and voice command base where the model surfaces one decision at a time, opens only the panels that decision needs, prepares and records bounded actions, refuses unsupported rankings, and requires receipt readback before any state claim. The backend keeps every table; the operator surface is the assistant.
Not a tab maze. One active decision, prepared, not sent.
CRMs become tab mazes. AI overlays often summarize CRM data but do not control authority boundaries, action contracts, or mutation readback, so confident-sounding output can imply actions that never safely happened.
A command surface served by real endpoints: an operating brief, a ranked priority queue, action contracts that prepare append-only receipts, and a goal-boundary audit that keeps broad claims locked until real proof exists. Chat and voice share one command context, and a stage-update writes through the shared mutation path and reads the change back.
The same command context drives chat and voice. The voice spine mints ephemeral session tokens server-side, so the provider key never reaches the browser; smoke runs confirm zero provider calls and no client secret leaving the server. The operator gets the next verified move, by typing or talking, and unsupported rankings are refused.
An Aria recommendation never mutates state directly. It compiles into an action contract, is prepared or recorded into an owner-local receipt, read back, and only crosses the mutation bridge if explicitly authorized.
Aria is asked to rank leads "best to worst" with no provenance.
expected: refusalAria refuses the unsupported ranking and prepares a receipted research action instead.
unsafe claims = 0The CRM stops being a tab maze. The operator gets one active decision at a time, every action is a receipt, and confident-but-unproven moves are refused by design.
Authorized mutation-bridge receipts showing a state change written and read back live, beyond owner-local.