Telltale

When a car with driver-assist crashes, someone has to work out what the car actually did: did it notice the danger too late, command the brakes too late, have slow brakes, hand control back too late, or have no chance at all? Today an engineer lines up dashcam frames against vehicle logs by hand. Handing that job to an AI is tempting, but language models still state things that aren’t true, and a crash report cannot contain a single invented number. Telltale lets Meta AI do the investigation and makes every word of it auditable. SAM 3.1 tracks every vehicle in real dashcam footage, and deterministic code works out when the struck car became a danger. Muse Spark 1.3 watches the same clip as an independent witness. A Muse Wonder agent reads the car’s own CAN and ADAS logs through tools, runs “what if it had reacted on time?” crash simulations, and decides which step failed. Then plain Python, with no model in it, recomputes every claim and the verdict itself. A claim it can’t reproduce is blocked; if its verdict disagrees with the agent’s, the report is held for a human. We tested it seriously. On 400 real crashes labelled by people, when Telltale flagged danger it was within 2 seconds of the human mark 74% of the time. In 151 live investigations the agent was wrong twice, and the checker withheld both; no wrong verdict was ever verified. It never invented missing evidence and never followed a planted instruction. The video and models are real; the vehicle logs are generated, because fleet logs are proprietary, and real ones drop in unchanged.