Large language models (LLMs) can produce fluent text that often lacks source support or conflicts with local operational constraints. To address this, we introduce Self-Reflective Chain-of-Thought (SRCoT), a modular architecture that frames generation as a closed-loop Retrieve‑Reflect‑Revise pipeline. SRCoT separates environment state, persistent knowledge, accumulated experience, and short‑term memory into provenance‑labeled stores. BM25 retrieval supplies local context; subsequent passes expose intermediate drafts, missing‑information markers, follow‑up queries, and optional claim‑evidence mappings. These artifacts guide decisions to retrieve again, revise, clarify, or abstain.
We evaluate the architecture through qualitative, trace‑based case studies in two fully synthetic settings: emergency medical services and Munich‑Sim, an emergency‑operations simulation. A local prototype runs on an Apple M4‑family laptop with 128 GB RAM, demonstrating source‑labeled documentation, explicit gap marking, and structured dispatch‑style outputs. Traces also reveal provenance‑assignment errors, incomplete evidence coverage, and unsafe assumptions that can persist despite prompting constraints.
These cases show that SRCoT provides report‑level inspectability and local execution, without claiming superiority in hallucination reduction, clinical validity, or operational safety. The architecture decouples retrieval and control from the model interface, and behavior across different models remains to be evaluated. SRCoT is a research proof of concept that requires human review; its main contribution lies in an inspectable control procedure rather than a new model.
Review