CairoFlow

CairoFlow is decision support for traffic engineers that shows its working. It turns one recorded overhead traffic video into tracked vehicles, turning-movement counts, speeds and a 3D digital twin of the junction, then refuses to publish any headline figure that fails its check. The problem: camera traffic analytics arrive as confident numbers with no way to tell which ones hold up, and engineers retime signals on them anyway. How it works: a detector and tracker follow every vehicle, SAM 3.1 refines the tracks and gives per-vehicle masks, a fitted camera placement puts them on the road, and analytics compute turning movements, speeds and stopped time. Before a single held-out read, we committed limits for seven checks against the published NGSIM trajectories, scored on alternating 60-second blocks the pipeline was never tuned on. On Lankershim Blvd at Universal Hollywood Dr (Los Angeles, 2005), 6 of 7 checks passed: recall 0.761, 0.863 of track rows matched to a published vehicle, 1.39 m across-road error (90th percentile), lane agreement 0.881, turning agreement 0.847, median speed error 0.48 m/s. Entering counts missed by 0.0033 (0.2033 against a limit of 0.20). We didn't move the limit, so the app withholds entering flow and shows the reason in its place. Muse Spark reads the checked evidence and writes hypotheses for an engineer to test, and it proposes signal-timing changes. Deterministic validators and the SUMO simulator test three candidates over 20 paired seeds across four signal timelines. Every candidate made something worse and no baseline is validated, so CairoFlow recommends nothing and says why. The product is a web console (Command, Vision Lab, Digital Twin, Analytics, Evidence) where every figure carries an evidence label and a gate status, all on one playback clock. It replays offline from cached model replies, with no API key, and has about 1,800 automated tests.