Closed-Loop Refinement of Generative AI Using Claim Decomposition and a Hierarchical Evidence Ledger
Abstract
Large language models operating in multi-turn agentic workflows can generate unverified statements or struggle to resolve conflicting information across disparate session sources. This disclosure describes systems and methods for a closed-loop refinement architecture to address these challenges. A system can programmatically decompose generated text into atomic claims and verify them against an evidence ledger. This ledger can organize session-specific information and use a configurable source-of-truth hierarchy to resolve potential conflicts. Based on the verification, a machine-readable correction prompt can be generated and sent to the model, enabling automated self-correction. This process can improve the verifiability and factual grounding of the output, allowing for granular source attribution for each statement.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Biswas, Susanta, "Closed-Loop Refinement of Generative AI Using Claim Decomposition and a Hierarchical Evidence Ledger", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12045