A traffic animation is not a detector test

A traffic simulator knows exactly where each authored vehicle is. It can draw a car, print its identity and show the route it follows. A roadside camera knows none of those things directly. It supplies pixels. The detector must find vehicles, the tracker must maintain identities, calibration must estimate road positions and the policy must retain enough history to justify an event. Both paths can produce a beautiful overlay. Only one has gone through the observation problem.

LaneSplitter's repository deliberately keeps these paths separate. The browser viewer consumes known trajectories and meshes. The roadside node consumes images, calibration and configuration. The evaluator can read authored truth to score its result, but the truth is not secretly fed into the detection pipeline. If we blur that boundary in a demonstration, the audience cannot tell which part of the system has actually been tested.

An authored-coordinate case is useful. It can check whether a rule distinguishes an ordinary merge from a prohibited queue cut when identities and positions are known. It can exercise permission boundaries, missing continuity and negative controls. The retained suite contains 82 such authored scenarios. Those cases give evidence about policy behavior under their inputs. They do not establish that a real camera can see every actor or recover the same movement history.

That distinction becomes important in crowded traffic. A vehicle may first become visible near re-entry, after the passing maneuver has already happened behind another row. The policy can receive a clean-looking track that lacks the decisive earlier observations. Lowering a confidence threshold cannot recover a history the camera never captured. The problem is coverage, geometry and continuity before it is a question of sensitivity.

The retained controlled boulevard render provides a bounded end-to-end example: camera frames pass through the real detector, tracker and production node, producing an expected event with saved evidence. A separate lawful-video control produces no event. These are valuable results because the input, expected behavior and saved output are identified. They establish what happened in those controlled scenes, not accuracy across a city.

The same review also retains misses. An earlier signed-divergence video did not produce its expected event. A later fixed-mast camera check processed real rendered pixels and produced zero flags, with early maneuver history not detected. We should show those outcomes next to their scope instead of collapsing all runs into a single pass percentage. A reader needs to know whether a new result addresses the same camera, behavior, configuration and visibility conditions.

A reusable result should identify its source scene, sensor viewpoint, time range, configuration and actual output. Presentation cameras are not necessarily inference cameras. A cinematic cut can make a maneuver obvious to a viewer while the fixed sensor had no usable view of it. The website therefore distinguishes a presentation of the scene from the measured sensor input. Moving between them should change the provenance shown on screen, not silently imply that they are interchangeable.

Identity is equally important for time. A slider that replays a different frame under the same timestamp makes comparison unreliable. When a retained moment expires or a required artifact is absent, the correct response is an unavailable state. It is better to state that a diagnostic was not recorded than to animate plausible policy stages. An explanatory model can be invented honestly; a result attributed to a real run must come from that run.

The useful next experiment is not always a larger, more dramatic simulation. It can be one difficult fixed-camera case with known permissions, complete history and a lawful control. Freeze those inputs, run the complete pixel pipeline and account for expected events, extra events and withheld candidates. If visibility fails, inspect camera geometry. If identity breaks, investigate tracking. If the evidence exists but the rule abstains, inspect the configured prerequisites.

The point of a rich simulation website is to make those distinctions easier to inspect. Good graphics can reveal cause and effect, preserve context and let a reader explore a decision. They become misleading when visual certainty exceeds measured evidence. A careful demonstration should leave the reader understanding both what the system did and what remains to be established.