Realist evaluations aim to evaluate interventions by understanding the mechanisms they trigger, assessing not merely what works but what works for whom, under what conditions, and how. Significant disagreement in the literature exists as to whether randomized trials can be used as a tool for realist evaluation. INCLUSIVE, which was the first realist randomized trial explicitly designed as such, evaluated the impact of Learning Together, a school-based intervention for students aged 12-15 that included restorative practice, on bullying victimisation, mental wellbeing and psychological problems. Drawing on hypotheses generated through qualitative research, this analysis tested if school belonging was a mediator of intervention effects, and in which contexts. We estimated a series of fully longitudinal multilevel moderated mediation models including intervention allocation, student reports of school belonging at 24 months and victimisation and wellbeing outcomes at 36 months, and stratified on the basis of whether, at baseline, schools were: a) rated 'outstanding' for leadership, b) below the median for average levels of victimisation, and c) above the median on school inclusivity. Findings suggested that in unstratified models, belonging was not a mediator for any outcome; but in each of the strata defined above, belonging was a significant mediator at the student level. However, in the strata where belonging was not a mediator, the intervention still had a significant effect on each outcome. Analyses point to a strong but conditional role for belonging as a mediator of intervention pathways; in schools where belonging was not a mediator (e.g. above-median victimisation levels), other mechanisms may have been activated. This is consistent with a realist understanding of context-mechanism linkages generating outcomes. Our analyses suggest that realist evaluations can be pursued within randomized trials and that such analyses can offer more nuanced evidence regarding in which contexts interventions might effectively be implemented.