Political forums have long been battlegrounds of misinformation where evidence is often drowned out by emotional rhetoric. For users of civic participation platforms, the experience is frequently a cycle of exhaustive arguments over claims that are never actually verified. This environment creates a vacuum where fake news thrives, not because of a lack of information, but because of the sheer difficulty of separating a subjective value judgment from a verifiable fact in real-time.
The Architecture of Asynchronous Verification
To combat this, the civic participation service Eoheung has deployed a sophisticated AI fact-checking pipeline designed for stability and cost-efficiency. The system does not process requests instantaneously; instead, it utilizes a Redis Cluster and BullMQ to manage a distributed in-memory data store and message queue. By pushing every user request into an asynchronous queue, the platform ensures that background workers process tasks sequentially. This prevents server crashes during traffic spikes and maintains a consistent user experience even when the volume of verification requests surges.
Model orchestration is handled through a tiered hierarchy to balance latency and precision. Gemini 3.5 Flash serves as the primary engine, handling the bulk of requests to keep response times low and operational costs manageable. However, the system includes a built-in fallback mechanism: if the primary model fails a complex logical inference or exhibits a high error rate, the request is automatically routed to Gemini 3.1 Pro. This tiered approach allows the service to reserve expensive, high-reasoning compute for only the most difficult verification tasks.
To prevent the API costs from spiraling, Eoheung implemented a strict resource management policy. Users are granted one free fact-check per day, after which additional requests require the consumption of in-app points. This economic barrier suppresses mindless API calls and ensures the sustainability of the infrastructure.
Filtering Noise to Eliminate Hallucinations
The primary challenge in automated fact-checking is not the lack of data, but the presence of noise. When an LLM is asked to verify a post that mixes subjective opinions with factual claims, it often attempts to validate the value judgment itself, which inevitably leads to hallucinations. For instance, a statement like "This policy is the worst in history" is a matter of opinion, whereas "This bill increases taxes by 20%" is a verifiable claim. Eoheung solves this by implementing a pre-processing layer that ignores value judgments and extracts only objectively verifiable factual claims before the verification pipeline even begins.
This process is enhanced by a multimodal context window. Rather than analyzing a single snippet of text, the AI is provided with the original post, parent comments, and attached images. By bundling these elements, the model gains a holistic understanding of the context, reducing the likelihood of misinterpreting a sarcastic remark or a nuanced argument as a factual claim. To further suppress hallucinations, the system is explicitly prompted to respond with "Unknown" or "Judgment Deferred" when authoritative evidence is missing, rather than attempting to force a conclusion.
Data integrity is managed through a snapshot system. Because community posts are frequently edited or deleted after a fact-check is requested, the system saves a permanent snapshot of the text and images at the exact moment of the request. This ensures that the verification result remains anchored to the specific version of the content that was analyzed, providing a transparent audit trail of the sources used and the evidence cited.
Success in LLM-based verification is not a product of model scale, but of architectural rigor. By prioritizing the extraction of verifiable claims and securing data integrity through snapshots, the system transforms the LLM from a creative writer into a disciplined auditor.
