How it's built, in brief
An honest overview of the architecture, with one stated limit: we don't describe the specific models and heuristics with which we extract and judge the claims. Not out of vanity, but so as not to help anyone who might want to game the system — the robustness of an automated fact-check also depends on not making it easy to fool.
Watch a fact-check happen
Pick a claim spoken live and follow the four steps, from audio to the outcome with its source. No real data: an illustrative reconstruction.
The pipeline
A broadcast's path is linear: we capture the audio of the public stream, we transcribe it, isolate the verifiable claims, compare them against the sources and publish the verdict on the real-time timeline. Each step is an isolated component: this makes the system robust and lets us improve one without touching the others.
Local transcription
Audio is transcribed on our servers. For extraction and fact-checking, text may be processed by external language-model providers, as described in the Privacy Policy.
Verification anchored to sources
Selected claims are compared with a curated primary-source index and, when needed, searches on authoritative domains. Verdicts include available sources, links and excerpts. Insufficient evidence leads to an unverifiable verdict. The explanation helps assess the comparison and its limits.
Real time and scale
The verdicts reach the browser via WebSocket and are synchronised with the player. The verification work is done only once per broadcast and streamed to every viewer: the system handles many people on the same event without multiplying the computation.
Infrastructure and data in Europe
Core infrastructure and account data are hosted on servers in Germany. This does not exclude processing by the external providers listed in the Privacy Policy. We use technical cookies and, with consent, analytics cookies; no advertising cookies. Revolut handles card data. Details in Privacy and Cookies.
Reliability
Each component has a dedicated health check and the processing jobs shut themselves down at the end of the broadcast: no resources left on for no reason, no orphaned jobs.
The areas we're working on are described on the Research page; the milestones already reached in Improvements.