Every claim passes through the same five stages, and each stage ships its evidence, so you can verify the verification.
Lenz takes a claim — one your product generated, or one you typed, as an API call in your workflow or a single pasted claim — and runs it through a structured pipeline: researched across multiple independent sources, scored for truthfulness, checked for bias, and backed by cited evidence you can audit.
-
Stage 1
Framing
The claim is received and cleaned up. We strip away emotional language and bias, distill the core factual statement, and prepare targeted search queries — so the rest of the process starts from a neutral, testable hypothesis.
-
Stage 2
Research
We search the web in parallel across multiple queries and independent search engines, then read the pages themselves. Quotes are extracted from each source's own text. Each piece of evidence is scored for authority, relevance, and recency.
-
Stage 3
Debate
We use two separate AI models to argue opposing sides in two rounds. First, one builds the strongest case that the claim is true while the other constructs the most compelling argument against it. Then each reads the opponent’s argument and writes a targeted rebuttal, exposing weak points in reasoning or evidence. Both draw exclusively from the collected evidence, ensuring every angle is stress-tested before the conclusion is reached.
-
Stage 4
Panel Review
Three independent reviewers, each assigned a model from a different vendor, read the same evidence and debate arguments and run the same three checks: how reliable and independent the sources are, whether the evidence bears directly on the claim, and whether the claim's numbers and wording match what the evidence actually supports. When enough quotes have been checked against their pages, one of them, chosen per claim, reads only those. Each scores the claim and explains its reasoning, without seeing the others’ votes.
-
Stage 5
Conclusion
All debates and analyses are synthesized into a single clear conclusion — True, Mostly True, Mixed, Mostly False, or False — with a Lenz Score from 1 to 10. A concise summary explains where the reviewers agreed or disagreed, and surfaces any important bias or logic warnings.
Ask follow-ups
Once a verification is complete, you can ask follow-up questions and get
answers grounded in the same evidence — the research brief, debate
arguments, and panel scores, plus additional sources pulled in on the fly.
Available in your product through the /ask endpoint, or by
chatting with Lenz right on any verified claim’s page on the site.
Why a pipeline, and not just ChatGPT
One model, asked once, is one opinion, sometimes from memory. Lenz implements an orchestrated combination of techniques that academic research has evaluated and shown to beat a single model, including:
- Independent research. An investigation of the evidence, run apart from the models: sources fetched, read and quoted in full, then weighted by authority, relevance and recency.
- Multiple models, multiple vendors. One model’s hallucination is another’s red flag.
- Adversarial debate. The strongest case for and against, argued on that evidence only.
- A panel vote. Three reviewers on models from different vendors, one reading only the quotes Lenz checked. Two more join when they disagree. Disagreements kept.
How much one confident answer hides: on 1,000 real claims, five frontier models failed to agree on 63%. Read the study →