Ask a single model and its answer might come straight from training, or from a web search on top. Either way you get an answer, not a verdict you can defend. Lenz is built to produce one:

• Grounded in retrieved sources. Lenz researches each claim from scratch — retrieving, scoring, and citing independent sources — so the evidence drives the verdict and travels with it.
• A panel, not a single voice. Models from different providers work each claim through the pipeline — arguing opposing sides, auditing each other. Different training data, different blind spots — one model’s hallucination is another’s red flag.
• A process built to surface the truth. Lenz argues each claim out — the strongest case for and against — then audits the evidence and arguments on separate axes: source reliability, logical fallacies, missing context. Every step comes back with the verdict, so you can check it, not just take it.

Today, the five frontier models, each ruling on the same 1,000 real claims, disagree more often than not: 67% of claims weren’t unanimous, 34% had a substantial split, and 21% had opposite verdicts (True vs False). That’s how much a single confident answer can hide. A panel surfaces the disagreement and settles it on the evidence. Read the study →

Evidence first, opinions never. As AI writes more of what people read, a confident wrong answer turns into a real liability — and there’s no transparent way to check AI output at scale. We’re building that layer: every claim held to the evidence, with every source and score on the record.