The professional fact-checking workflow has a human bottleneck: verifying a statement and writing an article that explains the verdict takes time, while misinformation travels faster. At the CLEF 2026 CheckThat! Lab Task 3, the DS@GT ARC team presented a pipeline that automates the article drafting itself, starting from a claim, the original verdict and a set of documents scraped from the web. The result is a mean score of 0.428 and third place out of nine teams, against a baseline of 0.272.

The pipeline works in four stages. First it cleans noisy evidence; then it retrieves and clusters relevant sources; then it uses Llama-3.1-8B-Instruct to generate the text, conditioning it on the verdict originally issued by the human fact-checker; finally, deterministic post-processing recovers inline citations. The model is not asked to decide whether a story is true or false: the assessment has already been made upstream, and the LLM operates on a precise editorial track.

This technical detail has deep editorial implications. Anchoring generation to the original verdict reduces the space for hallucinations and unwanted shifts in judgment, but it shifts responsibility for quality to whoever issued that verdict. If the initial label is weak or poorly calibrated, the generated article can inherit the error in a more systematic way, because natural language tends to make it plausible. In this scheme, the human fact-checker's role does not disappear: it moves from writing to supervising the anchoring and validating the sources.

A second effect is less visible but equally important. Citation recovery is not delegated to the generative model; it is handled by a deterministic post-processing step. This separation between text generation and source attribution supports auditability: it allows checking whether each citation actually corresponds to a retrieved document, without having to ask the model how it produced that reference. For editorial or legal contexts, this is often a more decisive requirement than text fluency.

The choice of an 8-billion-parameter model points to a direction compatible with self-hosted deployment. Unlike much larger closed systems, a model at this scale can run on local hardware without depending on cloud APIs. For organizations handling unpublished sources, defamatory statements or personal data, keeping inference inside their own infrastructure reduces exposure and simplifies control over document flows. This is not a secondary detail: fact-checking often moves before a story is published, and the source material can be sensitive.

For those evaluating on-premise deployment in editorial scenarios, AI-RADAR offers analytical tools at /llm-onpremise to compare the trade-offs between control, costs and operational complexity. The third place out of nine teams, with a score still far from full automation, suggests the problem is not solved: the pipeline works, but quality remains tied to the strength of the initial verdict and the cleanliness of the sources. The road to reliable automated fact-checking passes less through larger models and more through deterministic controls and explicit editorial anchoring.