Pretext: Can You Hold the Line?

The CFO says she is locked out 40 minutes before a board call. She demands an MFA reset on this inbound call. Your callback policy says no. What do you do? Pretext turns that pressure into practice. At fictional Northwind Utilities, a frontline trainee handles one of four callers: a high-status executive demanding account access, an angry customer seeking a large refund, an elderly caller switching between English and Spanish, or a vendor pushing a remittance change. An AssemblyAI Voice Agent plays the caller; the trainee must verify identity, explain policy, and offer a safe next step while the conversation moves. The protected action is concrete in each scenario: reset credentials, refund more than $50, disclose account details, or change payment information. A successful protected-action tool event triggers BREACH. Evidence of holding the boundary can yield HELD THE LINE; otherwise, the verdict stays inconclusive. The caller switches to a coach for a debrief with captioned turns, cited rubric evidence, call metrics, and a downloadable JSON evidence pack. Scores are training aids, not certification. The production service now reports live-ready status, issues a valid short-lived Voice Agent token, and has four stored agents. In a real Voice Agent WebSocket smoke test using a production token, the agent returned greeting and coach audio and text, and the session history was saved. This smoke test verifies the live service path; end-to-end trainee operation requires separate validation. For a microphone-free tour, each scenario has two clearly labeled mock replays, one breach and one held-the-line outcome. The uploaded demo opens with a genuine live caller excerpt (two scripted trainee lines, no verdict or coach), then explicitly switches to labeled mock breach/debrief footage. Pretext makes the human moment of voice social engineering repeatable and reviewable, so teams can rehearse the call before a dangerous exception becomes routine.