Every call leaves a test.

We use Coral when a private-pilot voice output has no saved voice choice. We keep your existing choice when you have one. This changes the fallback voice; we have not established a response-time improvement from the change.

Each new call now carries the carrier’s call identity through the conversation record, latency evidence, connection events and review receipt. A quick redial is therefore a new call even when it happens seconds later, rather than two conversations being inferred from a gap in timestamps.

Reception, calls SAINT places and hold rescue use the same envelope. Their outcomes are recorded in a bounded vocabulary: resolved, waiting, needs the person, failed or unclear. Unclear remains a valid result; the system does not manufacture resolution to improve its own record.

The review joins what SAINT said to the tool calls, ledger rows, confirmations, delivery state and connection events that could make it true. A supported completion can be marked clean. An unsupported claim, duplicate action, failed tool, long silence, repeated request, cut-off turn or connection error becomes evidence for review.

A confirmed defect can become a regression scenario or a proposed lesson about an existing tool. Neither route changes authority. Code, deployment, confirmation rules and fleet lessons remain subject to review.

The voice path now waits briefly before it cancels work after a possible interruption. It carries an unheard request into the next turn when the caller continues speaking. After a confirmation, SAINT reports completion only from the action executor's returned receipt.

We are testing repairs that keep a spoken approval attached to the proposal you heard. When an action is queued, we say so and wait for its execution receipt before calling it complete. We are also testing unfinished phrases, explicit booking confirmations and a single check-in when a call goes quiet. These repairs remain under validation for the private pilot.

We are also validating dinner coordination across text and voice. We keep your recent text request with the call, reject booking claims after a refused action, and read an old request against its original date. When you tell us you handled a booking, we resolve that booking while keeping any unsent invitation open. We look for address candidates in readable sent and received mail before asking you to repeat an address. These repairs remain in development, and we still need live pilot calls to verify them.

We are validating a further repair to how SAINT listens on calls. When you ask for silence, the call controller keeps brief acknowledgements quiet until you resume with a request. We also retain more of the current call's context and check completion claims before speech reaches playback.

We are tightening which messages can become your deadlines or interrupt you with a call. Our review now flags specific interaction failures even when the connection worked. These changes remain under validation; we still need live calls to establish their effect on listening and response time.

The per-call receipt stores the call identity, type, bounded outcome, review state, evidence counts and timestamps. It does not copy the transcript, request, summary, caller number, tool inputs or tool outputs.

Operator review now begins from that content-free receipt rather than loading the member's recent conversation. If an investigation genuinely needs the words from one call, the requester must state the reason, name a separate approver and use a time-bounded grant; the read is counted and remains due for disclosure to the member.

The receipt expires on the same ninety-day clock as the conversation and is removed with the account. This is private-pilot operational infrastructure, not a claim that every human quality of a call can be scored automatically. Warmth, speech accuracy and whether the person felt the matter was genuinely resolved still require consented human testing.