Skip to content

LenzVerification infrastructure for AI output

Audit-grade fact-checking for AI-drafted text.

Lenz checks the factual claims in AI-written text against independent sources before a customer or a client reads them, and flags what is uncertain.

  • €10,000 warrantyper qualifying check
  • Independent public sourcesnever the model that wrote it
  • Multiple vendors, three reviewersafter two models argue both sides
A check, end to end 81.6 s
“Spotify launched in the United States in July 2009.”
  1. 01 Framing claim, queries · 3.5 s
  2. 02 Research 18 public sources · 21.8 s
  3. 03 Debate for · against · 18.3 s
  4. 04 Panel Review 3 reviewers · 28.8 s
  5. 05 Conclusion verdict, score · 8.6 s
Verdict
False 1/10 high confidence
Suggested rewrite

Spotify launched in the United States in July 2011.

Sources · 18
  • techcrunch.com
  • theverge.com
  • engadget.com
  • +15 more

Each stage’s time is this run’s own, replayed faster. A full verification takes ~90 seconds per statement.

Open the full verification

01Why independent

A model grading its own output is not an outside check.

Lenz sits outside the model vendor and outside your own team. In a full verification it checks each factual statement against public evidence the model was never given: models from several vendors argue both sides and three reviewers weigh the arguments.

A split panel stays visible in the result. The evidence is public, so a customer or an auditor can follow it.

02How it fits

Lenz supplies the verdicts.
Your gate applies your thresholds.

  1. An answer is drafted

    A support reply, a RAG answer, a summary: anything a model wrote that someone is about to read.

    your model

  2. Statements are extracted

    The text is split into single, self-contained factual statements. Opinions, predictions and questions are left out.

    /extractno credits

  3. Assessed in bulk

    A panel of models returns a verdict and a confidence level per statement, synchronously.

    /assess~15 s

  4. Uncertain claims are investigated

    Negative and low-confidence results escalate to a full verification against independent public sources, with the arguments and reasoning.

    /verify~90 s · webhook

  5. Your gate decides

    What clears your thresholds ships. What doesn't goes to your reviewer with the verdict, sources and reasoning attached. Lenz never holds, edits or publishes your text.

    your policy

03Research · Open evaluation

Five frontier models, 997 claims, not unanimous on 63%.

Our open evaluation of LLM agreement on real fact-checks: frozen snapshots, full data, full methodology.

Read the report

  • 632 claims · not unanimous
  • 365 claims · unanimous
  • n = 997 · one dot per claim

04Warranty

What other AI products tell you: AI can make mistakes. Check important information.

The risk stays with you

Lenz warrants the verdict.

On a paid Pro or Scale plan, every full verification that qualifies carries a contractual warranty, at no extra charge. €10,000 per certificate, €500,000 across all Lenz certificates in any rolling 12 months.

The main conditions
  1. You published a warranted statement, or sent it to someone outside your organisation, after its qualified timestamp.
  2. A court or regulator later made a finding on the warranted statement that contradicted the verdict, once the time to appeal has passed or appeals are exhausted.
  3. Damages were awarded against you on that basis.

Fines, penalties, settlements and defence costs are not part of the warranty.

How the warranty works

A real warranty certificateTerms v1

The Great Wall of China is visible from the Moon with the naked eye.

verdict
False · 1/10
confidence
9.7/10
sources
29, from 26 domains
cover
€10,000 per certificate
certificate
fd7dcc7f…566f6deb
issued
qualified timestamp
Namirial,
OpenTimestamps
Bitcoin attestation
record hash
3f45d7b2…3e7f1163
Warranted verdictSee the certificate

05For developers

Four calls and one to run them all.

EndpointWhat it doesTime
/extractPull verifiable claims out of any textfree
/assessFast multi-model verdict for sync UX~15 s
/verifyFull adversarial pipeline with citations~90 s
/citecheckDoes each cited source say what the text says—
/reviewIssues in a draft's claims and its citations—

/ask answers follow-up questions about any verification.

API reference Integrations Connect Claude Connect ChatGPT

# pip install lenz-io
from lenz_io import Lenz

client = Lenz(api_key="lenz_...")
extracted = client.extract(text=model_output)
# identified_claims is empty when the text yields a
# single statement
claims = extracted.identified_claims or [extracted.claim]
r = client.assess(claims=claims)

# "Error" means not checked: never ship it as clean
for c in r.claims:
    if (
        c.verdict in ("False", "Mostly False", "Mixed", "Error")
        or c.confidence == "low"
    ):
        print("HOLD:", c.claim, c.verdict, c.confidence)
# HOLD: Python 3.12 runs about 20% faster … False high

Runs inClaudeChatGPTCursorClaude Coden8nZapierMCP serverRESTPython · TypeScript SDKs

Check the next draft before anyone reads it.

Free · 100 assessments or 10 full verifications a month