Lanesplitter. Keep the evidence close.

LaneSplitter investigates repeated queue cutting, informal extra rows and prohibited pavement bypass in Saudi-inspired road configurations. A single image of a vehicle between lanes is not enough to explain those behaviors. The system needs observations over time: where the queue was, whether the vehicle passed it, where it re-entered and whether it continued into the configured receiving region.

Road geometry and permissions are configuration, not something inferred from a familiar-looking sign. A split sign does not establish capacity or merge permission. The project therefore treats the site definition, calibration and rule prerequisites as explicit inputs. The work is to make a supportable observation, not infer a driver's intentions.

The roadside node captures camera or video input, runs the configured detector, maintains track identities and estimates road observations through calibration. Configured policies accumulate the history needed for a maneuver decision. Missing measurements, implausible movement or lost continuity can cancel an unsupported candidate. The output is a local evidence record with source images, measurements and identities.

The pipeline is implemented in a Rust workspace with separate research and runtime components. Local inference and evidence creation form the edge workflow. A future fleet-wide control plane is a separate product question, not a hidden dependency required for the node to create its first record.

The controlled boulevard example runs rendered camera pixels through the real detector, tracker and ordinary node. Its retained 720-frame run produced one expected queue-cut event and no extra events. The saved event clip preserves 324 frames through confirmation. A separate lawful-video control produced no event. These results establish bounded behavior in identified synthetic scenes.

The broader review also includes 82 authored-coordinate policy cases. Those use known trajectories and answer questions about rule behavior. They do not validate detector or tracker performance across the same suite. The public replay below keeps the actual pixel evidence separate and provides a receipt identifying the selected run.

An earlier signed-divergence camera run missed its expected event. A later fixed-mast diagnostic processed 316 rendered frames and emitted zero production flags. The early maneuver history was not detected; tracks appeared near re-entry. That is an observation-coverage limitation, not evidence that the full traffic sequence was successfully understood.

The useful response is to inspect camera geometry, visibility, identities and continuity before tuning sensitivity. A visually impressive 3D replay cannot restore pixels the sensor never saw. Keeping the miss beside the successful controlled example makes the next engineering decision more precise.

The node stores evidence locally before remote delivery. If that write fails, it skips the send rather than creating a remote-only record. A separate uplink task sends stored events and retries eligible backlog. A notification is a fast path to already-persisted work, not the only copy of that work. Configured retry limits remain part of the contract.

This separation does not promise survival through every failure. Disk capacity, power loss and the cost of writing evidence have their own consequences. The project guide identifies runtime throughput and global storage quota work as separate needs. The delivery model demonstrates the architecture while making its own illustrative limits visible.

Evidence includes source context, rule and configuration identities, measurements and artifact hashes. The review tools preserve source time and make retained clips watchable. A hash helps compare artifact integrity; it is not a signature or trusted timestamp. Missing clip coverage and unrecorded diagnostics must remain visible instead of being replaced by plausible animation.

Normal detection has no standing setup listener. A temporary setup interface is explicitly enabled for field configuration and disabled for normal operation. The next milestone is a defensible difficult camera case with a lawful control and a reviewable evidence packet. Field accuracy and deployment readiness need their own validation.