Delta-V

After a road crash, it is often one person's word against another's, even when a camera saw everything. The footage rarely settles anything: cheap CCTV drops frames, has no scale, and nobody can measure from it. Sakshi (साक्षी, "witness") turns ordinary road video into measured speed and timing, each with an honest error range, and checks what people say against it. Meta's models read the scene. SAM 3.1 tracks every vehicle, person and traffic light with pixel masks. Muse Spark reads the burned-in CCTV clock to recover true frame times, and reads traffic lights. Muse Wonder extracts checkable claims in Hindi or English and writes the report, and a validator rejects any number the engine did not produce. Muse Voice Transcribe handles recorded statements. A deterministic physics engine does every sum. It uses a homography from pixels to metres, a Kalman filter with smoothing, Monte Carlo uncertainty and collision checks. On 16 simulated crashes with exact ground truth, it gave 0 wrong verdicts in 63 claims, found 15 of 15 impacts, and had an 8.5% median speed error. In a real-world test, a car driven at 20 km/h measured 20.2 km/h. Sakshi supports expert review and never states who is at fault.