When synthetic Delphi panels work: a graph-native methodology for participatory foresight under structural complexity

Foresight methodology has historically expanded one mode at a time — from supernatural divination through hermeneutic utopias, technical forecasting, emancipatory foresight, and participatory creative futures. The widespread adoption of large language models opens a sixth mode, synthetic foresight , in which participants and evidence bases are computationally generated rather than recruited. We position this mode as the sixth column of an established typology of approaches to the future, extending the participatory tradition. This paper empirically evaluates synthetic Delphi panels and characterises the conditions under which they produce reliable forecasts. We deployed a 60-agent panel — six LLM families across ten disciplinary personas — over a Neo4j knowledge graph of 25,823 circular economy publications, testing eight evidence-base configurations: no grounding, full-text and abstract literature, EU regulatory documents, pre-extracted weak signals, and two graph-derived vocabularies (bridge-term injection via class-based TF-IDF and bridge-author traversal). Rankings were validated against observed 2021–2024 publication growth. The panel converged under every condition (CV reductions of 81–91% across three rounds), matching Delphi benchmarks. However, convergence was independent of coverage of the field’s research communities, indicating that agreement among synthetic panellists is not a sufficient validity criterion. Performance was sensitive to the type of grounding : structured bibliometric context substantially outperformed text-based retrieval on coverage, and the two graph-derived configurations achieved the best balance of coverage and ranking accuracy. We close with a reproducible four-step audit-first, deploy-second protocol and design guidelines, supporting synthetic foresight as a viable extension of the participatory tradition contingent on the quality of the knowledge infrastructure.