Lenz 是面向 AI 工作流的事实核查 API，通过多模型辩论和独立来源检查，为高可靠性需求的产品提供可验证的答案。
热门评论
PH 用户
Kosta here, co-founder of Lenz. Many businesses ship AI-generated content to their customers. Some of those use cases could benefit from factual verification of that AI output. That's why we built Lenz, packaged it as an API/SDK, and made it available across multiple platforms (n8n, Zapier, MCP, CLI), so people can easily integrate it into their workflows. Lenz verdicts come with a full audit trail - sources, citations, reasoning, confidence.
How Lenz is different than just asking a model:
(1) separate evidence gathering step (with source ratings) that doesn't rely on the model's memory or retrieval capabilities
(2) multi-vendor, multi-model approach to address single-model biases
(3) multi-round adversarial debate to crystallize the strongest for/against arguments
(4) multi-model jury reviewing the evidence and the debates across multiple axes
Key API primitives: /extract - extracts the factual claims from a text; /assess - quick assessment of a claim; /verify - the full deep claim verification; /ask - follow-up post-verification questions.
We measured the level of disagreement between the individual frontier models: on 23% of real-world claims, they disagree significantly, which sets the floor of the error Lenz is built to address.
To try a Lenz verification via the UI: lenz.io/verify
More about the LLM disagreement research: lenz.io/research/llm-disagreement
To integrate Lenz: lenz.io/integrations
GitHub: github.com/lenzhq
Hope you find this useful. Let me know either way :)
PH 用户
Hi, Vicky here 👋 the non-tech co-founder of Lenz
AI has genuinely changed what I can do on my own (from ops to marketing to building alongside an engineering team without being one). I’m one of those people it has really opened doors for.
But there’s another side to this. More and more companies rely on AI to produce the actual content their customers see: reports, articles, recommendations, research, support replies, ... And "someone will check it before it goes out" stops working at scale.
That’s the part of Lenz that really resonates with me. For editorial and production teams, that means letting verified claims through, sending uncertain ones for human review, and keeping the sources behind every decision.
AI will let us produce vastly more. Verification has to scale with it.
Very excited to have Lenz out here today and would love to hear your use cases.
PH 用户
Hi, I'm Pavel, part of the Lenz team. I worked on the MCP server and the CLI amongst other things.
Quick note on why we believe this needs to exist: we gave the same 1,000 real claims from Lenz to 5 frontier models (Fable, GPT 5.6, Gemini 3.1 Pro, Sonar Deep Research, Grok 4.5), identical prompt, all with web search and thinking. All five agreed on only 37% of them, and on 23% the verdicts were 2+ steps apart on a 5-point scale. Surprisingly confidence was almost meaningless: 76% of all answers were self-rated at 9 or 10 out of 10. The figure below is that result. Each ring is a model, each spoke is a claim, grouped by agreement.
The two services I built: - MCP server at lenz.io/mcp. To connect see lenz.io/integrations/mcp-server or just hand it to your agent. - CLI: pipx install "lenz-io[cli]", then lenz verify "the Great Wall is visible from space"
Paper, data and prompts are open if you want to pick it apart: lenz.io/research/llm-disagreement/v1.1
Happy to answer anything. Let me know if you try the MCP.
PH 用户
Congrats on the Launch @kostaj This hits close to me, spent Years in Security Testing and the same problem shows up there: one Models blind spot becomes everyones blind spot if you don't cross check.
PH 用户
Congrats team..idea of having a jury review the evidence is quite interesting.
How Lenz is different than just asking a model:
(1) separate evidence gathering step (with source ratings) that doesn't rely on the model's memory or retrieval capabilities
(2) multi-vendor, multi-model approach to address single-model biases
(3) multi-round adversarial debate to crystallize the strongest for/against arguments
(4) multi-model jury reviewing the evidence and the debates across multiple axes
Key API primitives: /extract - extracts the factual claims from a text; /assess - quick assessment of a claim; /verify - the full deep claim verification; /ask - follow-up post-verification questions.
We measured the level of disagreement between the individual frontier models: on 23% of real-world claims, they disagree significantly, which sets the floor of the error Lenz is built to address.
To try a Lenz verification via the UI: lenz.io/verify
More about the LLM disagreement research: lenz.io/research/llm-disagreement
To integrate Lenz: lenz.io/integrations
GitHub: github.com/lenzhq
Hope you find this useful. Let me know either way :)
AI has genuinely changed what I can do on my own (from ops to marketing to building alongside an engineering team without being one). I’m one of those people it has really opened doors for.
But there’s another side to this. More and more companies rely on AI to produce the actual content their customers see: reports, articles, recommendations, research, support replies, ... And "someone will check it before it goes out" stops working at scale.
That’s the part of Lenz that really resonates with me. For editorial and production teams, that means letting verified claims through, sending uncertain ones for human review, and keeping the sources behind every decision.
AI will let us produce vastly more. Verification has to scale with it.
Very excited to have Lenz out here today and would love to hear your use cases.
Quick note on why we believe this needs to exist: we gave the same 1,000 real claims from Lenz to 5 frontier models (Fable, GPT 5.6, Gemini 3.1 Pro, Sonar Deep Research, Grok 4.5), identical prompt, all with web search and thinking. All five agreed on only 37% of them, and on 23% the verdicts were 2+ steps apart on a 5-point scale. Surprisingly confidence was almost meaningless: 76% of all answers were self-rated at 9 or 10 out of 10. The figure below is that result. Each ring is a model, each spoke is a claim, grouped by agreement.
The two services I built:
- MCP server at lenz.io/mcp. To connect see lenz.io/integrations/mcp-server or just hand it to your agent.
- CLI: pipx install "lenz-io[cli]", then lenz verify "the Great Wall is visible from space"
Paper, data and prompts are open if you want to pick it apart: lenz.io/research/llm-disagreement/v1.1
Happy to answer anything. Let me know if you try the MCP.
这非常有力地提醒我们：AI自信的回答不等于经过核实的答案。
真的很喜欢你们用 Lenz 在做的事情。随着AI越来越融入我们的工作、学习和决策，对可信核查的需求只会越来越大。
到这一步，我已经很开心能遇到一个偶尔会说“让我查一下”而不是自信地瞎编的AI了 😂