Catch AI hallucinations before they ship.

Hallucination detection against public sources.

Send Lenz an answer your model produced.

It pulls out the factual statements and checks each against independent public sources, using models from multiple vendors. On a full verification, models argue for and against each statement, three reviewers assess it, and the result ships with the trace.

A model grading its own output is not an outside check. Lenz sits outside the model vendor and outside your own team, and the evidence it cites is public, so a customer or an auditor can follow it.

Inside Cursor and Claude Code with one command and your key; from a script in Python or TypeScript; or through the REST API.

The API: two different checks, one sequence.

Lenz supplies the verdicts. Your gate applies your thresholds.

1. 1

   An answer is drafted

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

2. 2

   Extract the statements

   POST /extract splits the answer into single factual statements, up to 20, each self-contained enough to check on its own. Opinions, recommendations, predictions and questions are left out. It costs no credits, capped at 1,000 calls a day per account.

3. 3

   Sort them, then establish the ones that matter

   POST /assess is a judgement: a panel of models returns a verdict and a confidence level per statement, synchronously, about 10 seconds, up to 20 statements a call; no score, no sources. Send the negative and low-confidence ones to POST /verify: an investigation against independent public sources, about 90 seconds a statement, asynchronous, with a webhook.

4. ✓

   Your gate decides

   Your thresholds and your routing: what clears them ships, and what does not goes to your reviewer with the verdict, the sources and the reasoning already attached. Lenz returns results — it never holds, edits or publishes your text.

5. 5

   Ship, with the record

   A full verification keeps its trace: the statement as framed, every source with its date, the arguments on both sides, the reasoning. API submissions are private by default; visibility: "unlisted" returns a link you can pass on.

The sequence is extract → assess in bulk → verify the uncertain results → ask follow-up questions about a verification. Assess to sort; verify to establish.

Try it on your own output →

Fact-checking RAG answers.

Retrieval narrows what the model says; it does not make it true. Lenz checks the factual statements in the answer against public sources. Faithfulness to your own retrieved documents is a different check, and not this one.

Check LLM output with Python.

# model output in — the statements your policy holds back, out
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)   # one row per statement, same order

# your routing policy. "Error" means the item was 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 than Python 3.11 ... False high

Frequently asked questions

Not answered here? Talk to us →

Catch the next hallucination before a customer does.