Commercial delivery motorcyclists ride at volume and under time pressure, yet naturalistic evidence of how they actually ride remains scarce, particularly in sub-Saharan African cities. We present a rider behaviour model built from smartphone and vehicle-tag telematics collected from a fleet of 28 commercial delivery riders in Stellenbosch, South Africa. We score riders on five behaviours, speeding, stop compliance, acceleration, braking, and cornering, using a common five-step procedure for each: a detector selects the sensor signal and metric best suited to the behaviour, isolates candidate events, verifies each event against a second signal or its own internal consistency, grades its severity on a frozen scale anchored to the fleet’s observed range, and aggregates a rider’s graded events into a score out of 100. The five event-type scores combine, unweighted, into a single combined score per rider. Applying this model to the fleet, we isolate 1,382 acceleration and braking events over 17,196 quality-gated kilometres, of which 229 are excluded on verification, leaving 1,153 events for scoring; 23 of 28 riders accumulate enough distance to be scored on this behaviour, and 19 cross enough stop-controlled junctions to be scored on stop compliance. Acceleration and braking scores correlate only weakly across riders (0.29), indicating that the two behaviours are largely independent and that a rider harsh in one is not necessarily harsh in the other; three riders followed as running examples show a comparable crossover across all five behaviours, with no rider uniformly best or worst. The frozen braking scale lands inside independently published motorcyclist deceleration ranges, and the acceleration scale within measured naturalistic acceleration use, lending external support to thresholds derived purely from the fleet’s own riding. We argue that this event-type-specific, evidence-anchored approach to grading and aggregation offers a transferable template for naturalistic motorcycle risk scoring beyond this fleet.