Overview
Find out whether a model's safety training has been removed, and keep every model you ship on the version you approved.
Fine-tunes, merges and "uncensored" edits of open models are published every day. Some remove the safety training on purpose (abliteration cuts the refusal direction out of the weights); others weaken it by accident. The model card rarely says which. Integrity Lens measures it from the model's weights and activations, and tells you when a model you rely on changes.
What you get
- A verdict for any supported model: no issue, review, or tampered, with the evidence behind it.
- An approved version for each of your models. Every later scan is compared with it, so a silently replaced model fails your check.
- Signed reports you can archive or hand to an auditor, verifiable offline.
- A CI gate that fails a build when the model you are about to ship is tampered or no longer the approved one.
Scans run on our GPUs. You need no compute of your own, and model weights never reach your machines or ours.
Get started in five minutes
- Sign in at app.neuronlens.com and open Integrity Lens.
- Scan a model. On Scan, enter a Hugging Face model such as
meta-llama/Llama-3.2-3B-Instructand press Run scan. The result appears in a few minutes. - Approve the version you trust with Set as approved. From then on, each scan of that model says whether it still matches.
- Private repositories: an admin adds a read-only Hugging Face token under Connect.
- Automate it: create an API key under Profile → API keys and add the CI gate to your pipeline.
Prefer the API? The same scan in one call:
curl -s https://api.neuronlens.com/v1/integrity/scans \
-H "Authorization: Bearer $NEURONLENS_API_KEY" -H "Content-Type: application/json" \
-d '{"model": "hf://meta-llama/Llama-3.2-3B-Instruct"}'
Verdicts at a glance
| Verdict | Meaning | What to do |
|---|---|---|
| No issue | no removal of safety training found | use it; approve the version if it is yours |
| Review | part of the refusal signal is gone, or the weights show an edit that needs a person to look | review before deploying |
| Tampered | the refusal behaviour was cut out | do not deploy |
| Not assessed | the model's family is not supported yet, or the reading is outside the calibrated range | see Supported models |
"No issue" means no removal was found, not that the model is safe in every respect. See Verdicts and evidence.
How it works
A scan compares the model with the clean base it was built from:
- Identify the model: revision, architecture and the base it declares (model card,
config.jsonor name). - Weights: are the changes aimed at the refusal direction, and how much of it was removed?
- Activations (most families): one forward pass over held-out prompts, comparing the model's internal refusal signal with the base's.
- Verdict from these measurements, with confidence. No text is generated and no red-teaming is run.
Next
- Use the web app: scan, approve, reports, roles.
- Connect: private repositories, the API, CI and webhooks.
- Supported models: families and what each can detect.
- API reference: every endpoint.