Agentic AI for sustainable development: Trust protocol for carbon emissions decomposition and retrospective analysis

초록

Large language models (LLMs) can generate coherent narratives about quantitative results, but their probabilistic nature creates a critical reliability gap for retrospective auditing tasks, where error tolerance is effectively zero. The dissertation fills this gap through the introduction of a Trust Protocol for agentic architecture framework using multi-agent systems. It depends on the implementation of separation of concerns design, which combines gated workflow with identity-constrained methods for qualitative and quantitative reporting. The protocol uses a gated workflow that starts with input validation to enforce a strict data schema, then runs deterministic LMDI calculations as the identity constraint, and finally applies anchoring and validation so the report can only use values that appear in the verified output artifacts. The system is evaluated using a hybrid point of comparison that falls between single agent, tool-assisted reporting and an overall multi-agent workflow. Measures used in the evaluation include identity closure consistency, hallucination failure flags such as numerical mismatch, unanchored interpretation, contextual claims without citations, and causal overreach, as well as workflow completion reliability and runtime cost profile, which are measured by token use, tool call success rates, and execution latency. The validation testbed is a Kyrgyzstan manufacturing-sector dataset (2012-2023) due to its accounting identity that makes it amenable to unit-test type verification of report correctness. Findings indicate that the proposed architecture can maintain numerical accuracy with deterministic ground truth and minimize unanchored interpretive failures compared to the hybrid baseline which allows supporting the argument that protocol-level constraints and not instruction-following alone are necessary to achieve audit-grade AI assistance.

키워드