A synthetic city, a few hundred independent movers, and a simulated wide-area sensor with Corvus-style exploitation running live on top: detection, persistent tracking, backtracking. Click any mover to trace it back to its origin. Then turn the difficulty dials — density, sensor rate, jitter, occlusion, contrast — and watch the tracker degrade, honestly, measured against perfect ground truth. Or press Run benchmark for the reproducible test matrix (fixed seed: same numbers on every run in the same browser).
| Config | Movers | Sensor | Jitter | Occl. | Contrast | Detection | ID sw/min | Tracks (avg) |
|---|
Faint dots are the true scene, moving continuously. The brighter markers are what the exploitation layer knows: a sensor frame arrives 0.5–2 times per second, detections carry noise and jitter, occlusion swallows movers, and close passes merge into one detection — so tracks must survive gaps and clutter.
Click any mover: that is your event, and its history draws back to its origin with the time delta. Then starve the sensor to 0.5 fps, add occlusion, collapse the contrast — and watch the identity switches climb. The dials manufacture the failure cases that matter, on demand.
Every pixel is generated: no real persons, vehicles, or places, and no clearance needed. Detection rate and identity switches are measured against ground truth, live — and the benchmark runs on a fixed seed, so every improvement to the tracker lands as a before/after row, not an adjective.
FULLY SYNTHETIC — EVERY PIXEL GENERATED. SLICE 2 OF THE PUBLIC DEMONSTRATION: DIFFICULTY DIALS + REPRODUCIBLE BENCHMARK. THIS SLICE IS ARCHIVED AS SHIPPED; THE CURRENT SLICE LIVES AT /demo/.