Skip to content

How it works

How Lenz checks a document.

Give Lenz a draft, a document, or a single claim. It extracts the factual statements and the citations, checks each cited source and factual claim one by one, and suggests minimal edits where the evidence shows a fix — with sources and reasoning you can inspect.

  1. 01Extractstatements & citations paired
  2. 02Citationssources resolved & read
  3. 03Fact checksquick check, then deep check
  4. 04Editsminimal fixes, verified

01Extract

Pull out the factual statements and the citations.

Pass in text, a web URL, a Google Doc, a Word document, a PDF, or an image. Before checking anything, Lenz reads the document and builds two inventories so nothing is silently skipped.

  • Factual statements

    Lenz separates checkable assertions from surrounding commentary and ranks up to 100 statements with the most consequential claims first — figures, dates, study findings, and statutory references — while leaving opinions and forecasts alone.

  • Citations and references

    Every inline link, DOI, footnote, author-year reference, and legal citation is paired with the specific statement it is meant to support, so each reference is checked in the context where the draft uses it.

02Citation checks

Read each cited source and check whether it backs the sentence.

A link that loads is not enough. Citations fail when a reference does not exist, or when the page loads but does not say what the draft attributes to it. Lenz checks both.

  1. Resolve the reference

    Web links are fetched directly. Academic papers, DOIs, and author-year citations are looked up across scholarly registries, and legal citations are resolved against official statutory, regulatory, and case-law databases across the US, the EU, the UK, and Germany — alongside any reference documents you supply.

  2. Judge support against the text

    Lenz reads the retrieved source and checks whether it supports the paired statement — Supported, Partly supported, Contradicted, or Not in source — returning the exact passage that backs the finding. When a source sits behind a paywall or cannot be read in full, Lenz marks it Unchecked with the reason rather than guessing at unread pages.

  3. Check verbatim quotes

    When the cited sentence places words in quotation marks, Lenz checks that the quote appears in the source text and flags fabricated or altered quotations.

03Fact checks

A quick check across the draft, a deep check where it counts.

Running a ninety-second investigation on every background sentence in a long report wastes time and budget. Lenz verifies factual statements at two levels of depth and escalates the claims that need a full investigation.

  • Quick check

    about fifteen seconds

    A three-model panel across different AI vendors evaluates each statement in parallel and returns a quick verdict, a confidence level, a short note, and the pages the reviewers looked at. Across a document, quick checks separate statements that hold up from the ones that need a closer look.

  • Escalation

    where a closer look is needed

    Statements that come back disputed, uncertain, or contradicted are the ones that need sourced proof. In automated document reviews (the API, Google Docs, n8n, and Zapier), Lenz escalates those claims to a deep check automatically. In the Workbench and in Claude or ChatGPT, you can also run a deep check on any statement with one click.

  • Deep check

    about a minute or two

    Lenz investigates the claim across five stages: framing a neutral research plan, reading independent sources (or your own reference documents), arguing both sides in a two-round debate, scoring the evidence with a panel of reviewers from different vendors, and writing a sourced conclusion with a Lenz Score from 1 to 10.

The deep check can change the quick verdict. When it does, rely on the deep check: it is the one that shows its sources and its reasoning.

04The deep check

Five stages, one sourced verdict.

Whether a claim escalates from a document review or is sent directly as a single claim, the deep check runs it through the same five stages. Every stage ships its evidence and reasoning in the trace so you can verify the verification.

  1. Framing

    The claim is pinned exactly as written and a research plan is built around it: targeted search queries, the entities involved, and whether the claim is time-sensitive — so the rest of the process starts from a neutral, testable hypothesis.

  2. Research

    We search in parallel across multiple queries, independent search engines, and academic, legal, and financial registers — alongside any reference documents you supplied — 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.

  3. Debate

    Separate AI models 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, so each argument is challenged before the conclusion is reached.

  4. Panel Review

    Three independent reviewers read the same evidence and the same debate. Each runs on a model from a different vendor, and each applies the same three checks: are the sources reliable and independent, does the evidence bear directly on the claim, and do the claim’s numbers and wording match what the evidence supports?

    Nobody sees anyone else’s vote. Each scores the claim and explains the reasoning behind the score.

    The verdict is the median. When the panel splits, two more reviewers join.

  5. Conclusion

    The panel’s verdict — True, Mostly True, Mixed, Mostly False, or False — is written up 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. When no public record in the evidence settles the claim, Lenz answers “No Public Record” instead of a score. That doesn’t make the claim false.

05Suggested edits

Minimal fixes anchored to your draft, verified before you see them.

When a statement is wrong or overstated and the evidence establishes the correction, Lenz drafts the smallest edit that makes the sentence accurate.

  • Surgical by design

    Each edit targets the exact span in your draft — replacing a wrong date, figure, or statutory citation, or qualifying an unsupported absolute — while keeping your surrounding prose and voice intact.

  • Guarded against new errors

    Every candidate edit is verified before it is shown so a fix never alters a direct quotation, drops surrounding claims in the sentence, or introduces a new inaccuracy. If a candidate edit does not clear those checks, it is withheld.

  • Applied where you work

    Apply accepted fixes with one click in the Workbench or the Google Docs add-on, export an annotated Word document with reviewer comments and native Track Changes, or consume the character offsets and replacement strings directly from the API.

06Why a pipeline

Why a pipeline, and not just asking one model.

Pasting a document into a chat window and asking “check this for accuracy” feels like a review. When the document carries your name, a single model falls short in three ways — and why Lenz is built as an orchestrated system instead:

  • A single prompt skims the draft

    Ask one model to review a twenty-page report and it gives an editorial impression. It flags a few conspicuous sentences, skips most of the claims, and rarely opens every cited URL, DOI, or statute to verify what the source actually says. Lenz extracts the statements and citations first and checks them individually, so you see every item that was checked, what held up, and what could not be read.

  • Models make errors with high confidence

    One model, asked once, gives one opinion — often from memory — and states it with conviction even when it is wrong. Lenz separates source retrieval from judgment, argues both sides from the retrieved text in an adversarial debate, and takes the median vote of independent reviewers across different AI vendors, so one model’s hallucination is another’s red flag.

  • Disclaimers instead of accountability

    General AI assistants end with a disclaimer that AI can make mistakes, leaving the risk with the person signing off. Lenz returns the full source and reasoning trace so your team can inspect the work, and backs qualifying verifications with a financial warranty.

  • Specialized model roles, continuously evaluated

    No single model is best at claim extraction, legal citation resolution, adversarial debate, calibrated panel scoring, and surgical copy editing. Lenz assigns over thirty specialized model roles across five AI vendors, matching each task to the models that perform best on it, and continuously benchmarks models and refines prompts against internal evaluation suites before changes ship.

How much one confident answer hides: across 997 claims, five frontier models were not unanimous on 63%.