The Topology of Economic Fragmentation: Semantic Anti-Agglomeration and Structural Lock-ins in the South African Labour Market

Structural transformation, defined as the continuous relocation of labour and capital across economic sectors, remains the primary driver of long-run development and regional prosperity. In contemporary settings, multiple rapid technological transitions occur simultaneously, rendering broad macro-statistics increasingly insufficient for capturing the granular shifts in industrial evolution. To address this gap, recent literature in economic complexity characterises economies as networks where the proximity of industries captures shared productive capabilities, skills, and technologies. This paper operationalises a "genotypic" variant of this perspective by using Sentence-BERT transformers to compute pairwise semantic similarity directly from official industrial classification descriptions, creating a high-resolution "Semantic Space". Projecting sectoral employment data onto this semantic space allows us to view the labour market as networks at different levels of sectoral and spatial aggregation. While theory suggests that growth should propagate along semantically proximate edges through localised knowledge spillovers, cities of the global South face a structural transformation paradox: the sectors that grow do not appear to draw their nearest industrial neighbours along with them, even when those neighbours share deep operational and cognitive similarities. This paper examines that paradox empirically in the South African context, testing whether the near-zero semantic autocorrelation of labour market transitions observed at the national level is, as theory would predict, an artefact of spatial aggregation that dissolves when the analysis is disaggregated to the municipal and intra-city scales. We find that spatial disaggregation fails to surface the predicted positive autocorrelation, with the semantic Moran's I shifting toward more negative values-from +0.004 nationally to-0.035 at the intra-city scale. These results reveal a structural feature we term "semantic anti-agglomeration," where semantically proximate sectors systematically diverge in employment growth due to competitive substitution, structural lock-in, and extreme spatial fragmentation. Such findings demonstrate that in highly unequal urban systems, physical and semantic proximity are doubly decoupled: conceptually adjacent industries fail to co-locate, and even when they do, institutional barriers prevent them from translating that proximity into shared dynamics. Ultimately, this research implies that standard cluster-based industrial policies are unlikely to function without first addressing the underlying urban substrates-such as transport and housing-that allow economic geography to become economic opportunity.