How Fountem works
We don’t ask a chatbot for its opinion. We break a claim into checkable parts, retrieve primary-source evidence, and only then produce a verdict — one that always cites where it came from.
This page describes only what the system actually does today.
1. Atomic claim decomposition
Political statements often bundle several assertions (“X built more homes and cut taxes”). We split a claim into independent, individually checkable sub-claims. Each is verified separately, then aggregated — so a single false component can’t hide inside a broadly true sentence.
2. Retrieval over primary sources
For each sub-claim we run hybrid retrieval — semantic vector search plus keyword (BM25) search — over a curated corpus of primary sources: official statistics (ONS, government data), Hansard, manifestos and reputable reporting. We retrieve the most relevant passages rather than relying on a model’s memory.
3. Grounded, sourced verdicts
A language model is given only the retrieved evidence and asked to reach a verdict that cites it. Verdicts use a calibrated scale:
True / Mostly trueHalf trueMisleading / Mostly falseFalseUnverifiable
4. We'd rather say “unverifiable” than guess
If retrieval returns no adequate evidence, we return unverifiable instead of inventing an answer. Every verdict also ships a “what would change this” statement so you know exactly what new evidence would move it.
What we won't claim
We’re only as good as our corpus. A verdict reflects the evidence we have indexed; gaps produce “unverifiable”, not confident error.
We don’t replace human judgement. Verdicts are a sourced starting point for scrutiny, not the final word.
We show our working. Every claim links the exact sources and excerpts used so you can check us.