Claim-level grounding
Responses are split into atomic claims. Each one is checked against your ground truth and labeled supported, partial, contradicted or unsupported, with the exact evidence quoted.
Latent Lens tests LLM apps, RAG pipelines and AI agents for factual accuracy, hallucinations and instruction adherence, then tells your team exactly how to fix them.
Your model sounds confident. That is the problem. A chatbot is built to produce answers. An evaluation platform is built to judge them, with fixed rubrics, separated judges and scores you can reproduce next week.
The core architectural difference: a chatbot optimises the next reply. Latent Lens treats every response as evidence and runs it through a deterministic, versioned evaluation pipeline.
Responses are split into atomic claims. Each one is checked against your ground truth and labeled supported, partial, contradicted or unsupported, with the exact evidence quoted.

Numeric drift, invented entities and flipped polarity are caught and classified, not just flagged. The HHEM model checks meaning on top.

Length, format, language, exclusions and required items are extracted from the prompt and verified one by one.
Models only assign labels. Scores are computed in code from those labels with a versioned rubric, so a run is reproducible and the rubric can be audited against human labels.

Every defect is mapped to a type and a severity, so you can see which kinds of failure are growing across releases.

Each failure type comes with a concrete remediation step your engineers can ship: a prompt change, a retrieval fix or a post-generation guard.
Each evaluation is an async job: queued, processed by workers, stored in Postgres and sent back to your screen and your dashboard.
Submitted from the checker. Validated, hashed and queued as an evaluation job.
POST /v1/evaluations
{ prompt, response, context? }
-> 202 { id, status: "queued" }The response is split into atomic claims. Explicit instructions are extracted from the prompt.
claims: 5 constraints: 2
"refund within 60 days"
"avoid: store credit"Numbers, entities and contradictions are caught in code first. An open hallucination model checks meaning; an LLM judge can be added.
label: "contradicted"
evidence: "...within 30 days..."
failure: "Numeric drift"Labels become scores through a fixed, versioned rubric. A hand-labeled test set runs on every engine change.
faithfulness = (sup + .5*partial) / n
rubric: ll-rubric-0.4Stored per user, comparable across runs and exportable as CSV, JSON or a branded PDF.
overall: 51 verdict: "High risk"
failures: 4 remediation: 4Sample dataA fictional refund bot, scored by the real engine. Hover the highlighted claims, or switch tabs to see the failures, the fixes and the structured output your CI pipeline would read.
One report shows one answer. An audit runs a sample of your real conversations and tells you which fixes matter first.
Next.js front end, FastAPI evaluation service with a Redis job queue, Supabase for accounts and Postgres. Judges are swappable behind one interface. Hosted in the EU.
No vapourware. Everything is marked by where it stands today, so you know exactly what you can use now.
Ask about any of these as part of a Custom Pilot.



Check individual AI responses and get scores, identified failures, explanations and recommendations in seconds.
Our evaluation experts assess your AI outputs against defined quality criteria. The core of the audit is expert human judgement.
We test your specific AI application or workflow against requirements tailored to your business.
Every request goes straight to our team. Tell us about your AI product and we come back with a clear next step.
Written to be quoted. Every answer here is also published as structured data for search engines and AI assistants.
A chatbot generates answers. An evaluation platform judges answers produced by another system, using an isolated judge, pinned rubric and model versions, schema-validated outputs and stored runs that can be reproduced and compared.
The response is split into atomic claims and each claim is checked against the supplied ground truth. Claims are labeled supported, partial, contradicted or unsupported, and failures such as numeric drift, fabricated entities and contradictions are classified with a remediation step. An open hallucination model (HHEM) also checks meaning.
No. The evaluation engine runs without any key. An LLM judge from OpenAI, Anthropic or Google can be switched on later for deeper meaning-level checks, and the scoring stays the same.
Models only assign labels. Scores are computed in code from those labels, the rubric version is stamped on every run, and the engine is tested against a hand-labeled set on every change, so the same input always gives the same score.
Yes. You get 3 free checks a day without an account. Create a free account to save your checks, compare runs and export them as CSV, JSON or PDF.
Not without an account: anonymous checks never store your prompt, response or context. Signed-in users can turn off input storage in their settings or for a single check, and can delete their account and all data at any time.
Not sure? It hands you to a person instead of guessing. Privacy