Automated fact-checking with LLMs suffers from a trust problem: in an end-to-end prompt, evidence retrieval, reasoning, and uncertainty estimation get tangled together, making failures hard to diagnose. R2VC tries to break this knot with a modular pipeline that separates four stages: retrieval, candidate generation, verification, and calibration.

The system combines hybrid sparse+dense retrieval over Wikipedia, a supervised fine-tuned and DPO-aligned generator that produces diverse structured verdict candidates, an external NLI cross-encoder for evidence-based candidate selection, and a lightweight sequence-level calibrator that estimates confidence and decides when to abstain. This is a precise architectural choice: instead of asking a single pass to do everything, it builds separate, auditable components.

Results on FEVER give the design real substance. With an 8B backbone, R2VC improves accuracy by 13.74% over baseline. Ablations show that the largest contributions come from verifier-based candidate selection and calibration: removing selection drops FEVER accuracy to 76.24%, while removing calibration nearly doubles the Brier score to 0.161. The message is that external verification and calibration are not add-ons—they are structural parts of the pipeline.

A manual analysis of 250 errors adds a crucial detail: retrieval failures, especially wrong-entity evidence, remain the dominant bottleneck. This means even strong verification cannot compensate for weak retrieval. Those designing fact-checking systems need to invest first in the quality of retrieved sources, not only in downstream reasoning.

For anyone looking at architectures built for self-hosted environments and data sovereignty, the proposal has structural value: a modular pipeline makes it possible to intervene on individual stages, replace retrieval or the verifier without rebuilding the whole system, and keep local control of each component. The source does not specify deployment context or hardware requirements, but this separation of tasks is exactly the kind of design that supports on-premise evaluation, where transparency and auditability matter as much as accuracy. For those evaluating on-premise deployment of similar fact-checking pipelines, AI-RADAR offers analytical frameworks at /llm-onpremise to assess the trade-offs, without prescribing a single choice.

An open question remains whether the calibration gains hold on domains different from Wikipedia or on private corpora, where retrieval is often the weakest point. R2VC nonetheless points to a clear direction: trust in LLMs is not built with longer prompts, but with architectures that separate and control the sources of uncertainty.