Reality Fork

Reality Fork Film anything that moves. Get its physics. Then change them. Live app: https://reality-fork.onrender.com Whether a person hit by a car survives depends mostly on the car's speed, which is why the 2020 Stockholm Declaration on road safety asks for 30 km/h wherever people on foot mix with traffic. Parents who want that outside a school are asked for evidence: a speed survey, summarised by the 85th-percentile speed. That takes a radar gun or a council traffic count. Reality Fork does it from video. Muse Spark 1.3 acts as the scientist. It watches the clip, decides what to track and what sets the scale, and calls our tools through its native function calling. SAM 3.1 follows the car through every frame, and its masks become sub-pixel measurements. Seven laws of motion are fitted by Levenberg-Marquardt and ranked by BIC, with errors propagated from the scale. The agent then draws its law back over real frames and checks the fit before it reports. On a junction in Illinois it measured a car at 17 +/- 1 km/h. Then it forks reality. Someone steps out 20 m ahead: at its filmed speed the car stops 11 m short; at 30 km/h, 2.5 m short; at 40 it still hits at 32 km/h. The same lab is an MCP tool for coding agents: eight clips of one junction become a survey, with a mean of 30 km/h and an 85th percentile of 33 km/h, carrying a Monte Carlo error bar. It also works on any motion. From 1972 Apollo 16 TV footage it measured the Moon's gravity at 1.85 +/- 0.18 m/s^2 (truth 1.62), and Artemis I's acceleration within 2.5%. The median error in g across test scenes is 0.75%. Muse Image repaints scenes as the Moon, Mars or Jupiter for gravity forks. Built during the hackathon (27 Oct–4 Nov 2026): everything in the repository—agent, physics fitting, fork simulation, web app, MCP server and CLI, and tests. No earlier project code is used. Third-party pieces: Meta's @meta-sam/parser and ffmpeg. Test footage: NASA (public domain) and a Wikimedia Commons junction video by Illinois_Roadbuff, CC BY 4.0, credited in clips/SOURCES.md and the demo's closing card. Limits: the junction speeds have no radar reference, so they demonstrate the method rather than validate it; the oblique view and assumed car length add error. Braking assumes a 1.5 s reaction time and 7 m/s^2 braking; every summary states that.