Get in touch

Grounded models

Models that are grounded in your documents.

Every sentence is traced to its source.

Works on open-weight models.

Built by researchers from Fraunhofer HHI.

Demo

Every claim traced to its source.

Models add claims their sources do not contain. Layerwise links each claim to its source and flags claims it cannot trace or that are likely wrong.

Policy assistantExample of a bank
Checked by Layerwise
We want to move our payment processing to a new cloud provider. What do we need before signing, and when do we tell the regulator?
Payment processing is a critical function, so the Outsourcing Committee must first approve a risk assessment and an exit plan.1
Group Compliance then sends the notification to the supervisory authority.3
This must happen at least 30 days before the contract is signed.No source
2 of 3 sentences cited1 held for review
Click a sentence to see its source
Documents given to the model
Sentence 1 comes from document 1.
1Section 2.1 · Critical functions
Payment processing is a critical function. It may only be outsourced after the Outsourcing Committee has approved a risk assessment and an exit plan for the provider.
2Section 4.1 · Provider monitoring
After go-live, the business owner reviews the provider's service levels and reports serious incidents to the Outsourcing Committee.
3Section 5.2 · Supervisory notification
Group Compliance notifies the supervisory authority of every outsourcing of a critical function and sends the approved risk assessment with the notification.
Words the selected sentence relies on

Offer

Answers you can check against your documents.

Included in every deployment

Citations

Every sentence links to the exact words in your documents it came from.

  • A source for every sentence of the answer
  • The exact words in the document are marked

Warnings

Sentences without a source or likely to be wrong are flagged for review.

  • Flags answers that come from memory
  • Catches wrong details in sourced answers

Optional service

Grounding tuning

We tune your model to your documents, so it answers from them more reliably.

  • Tuned for your use case by our scientists
  • Can be switched off for any request

Proof

Results on open-weight models.

On hard benchmarks, where the document contradicts what the model learned. Measured on Qwen3.8-27B.

Included in every deployment

Wrong answers caught

Layerwisereads the model that writes the answer
91%
No extra model
LettuceDetect-v2dedicated detection model
55%
+ 307M model
HHEM-2.1dedicated detection model
45%
+ 250M model
TinyLettucededicated detection model
39%
+ 68M model

With 1 in 10 correct answers flagged for review.

Optional service

Answers that follow the document

Best written prompt vs. after tuning.

+64%50% → 82%

When the document contradicts the model

+50%60% → 90%

On unseen law and finance documents

Deployment

Deployment on your infrastructure.

Your infrastructure

It runs inside the serving of your model, on your servers or in your cloud.

Your data stays with you

Documents, questions and answers do not leave your environment.

Open weights, unchanged

The base model stays as published. The tuning can be switched off for any request.

Pilot

Eight weeks, one workflow, one model.

You get cited answers, warnings and a report for your risk committee, plus a tuned model if you choose. The report addresses what the EU AI Act asks of high-risk systems: logging (Art. 12), human oversight (Art. 14) and accuracy (Art. 15).

  1. Week 1

    Scope

    Choose the workflow and the model.

  2. Week 2

    Connect

    Use our endpoints with citations and warnings from day one.

  3. Weeks 3 to 6

    Test and report

    On your questions, then on real requests. You decide on rollout.

  4. Weeks 7 to 8

    Tune, optional

    We tune the model to your documents, so it answers from them more reliably.

  5. After week 8

    Ongoing support

    In production, and when your documents or your model change.

What we need from you

  • An open-weight model

    Served by you or by us.

  • Your documents and test questions

    To measure the result.

  • One engineer as a contact

    For access and the endpoint switch.

For your developers

The answer streams as before, with sources and warnings for each sentence.

Developer documentation

Test it on your documents.

Tell us which model you run and which documents it should answer from.

Get in touch