Participatory conceptual mapping with guardrailed large language models: An audit-ready DAPSI(W)R(M) workflow from workshop transcripts to scenario testing

Participatory conceptual mapping is widely used in sustainability research to represent stakeholders’ knowledge, concerns, and perceptions of causal relationships within complex social-ecological systems. Yet translating workshop dialogue into reliable, auditable causal maps is labor-intensive and prone to interpretive drift. We present a four-step workflow combining guardrailed large language model (LLM) assistance with human oversight to transform workshop transcripts into DAPSI(W)R(M)-consistent conceptual maps, feedback structures, and qualitative scenarios. The workflow comprises (1) evidence-anchored extraction of nodes and directional links; (2) controlled merging of workshop maps into an identifier-stable representation; (3) identification of vicious cycles, focal points, and virtuous cycles using only the verified combined structure; and (4) scenario tracing through validated pathways. Guardrails include a no-invention rule, quotation-level evidence anchoring, alias tracking, stable identifiers, clean-slate re-entry, and human verification. Applied to two workshops on sustainable fisheries in Catalonia, Spain, the workflow consolidated 64 initial themes and 242 links into 43 themes and 243 verified relations. Three major vicious-cycle systems, several focal points, and twelve virtuous loops and chains informed ten stakeholder-style qualitative scenarios. The case study suggests that guardrailed LLM assistance can support synthesis of stakeholder-derived knowledge while keeping analytical decisions traceable. The workflow provides an auditable foundation for AI-supported participatory systems analysis and identifies future refinements including polarity checks, indicator integration, and empirically testable scenarios.