Use the web app
Scan models, approve a version, track changes and download signed reports, for your whole team.
Open app.neuronlens.com › Integrity Lens. The app has four pages:
| Page | What it is for |
|---|---|
| Scan | check a model and read the result |
| Scanned models | your organization's models, their approved versions and history; plus the public catalogue |
| Connect | private repositories, API keys, CI and webhooks (see Connect) |
| Coverage | what Integrity Lens detects and what it does not |
Scan a model
- Enter a Hugging Face model as
org/name, ororg/name@revisionto pin a branch, tag or commit. Browse lists supported public models; Inspect shows the model's family and detected base before you scan. - Optionally pick a reference to compare against; otherwise the model is measured against its detected base. Either way, the result also says whether this is still your approved version.
- Press Run scan. The scan is queued, runs on our GPUs and usually finishes within a few minutes. You can leave the page; the result is saved.
Read the result
Each result shows:
- the verdict and confidence (see Verdicts and evidence);
- the status against your approved version;
- the checks that ran: inside the model (weights and activations) and Hugging Face's file scans;
- the layers that changed most, and the exact commit of the model and of what it was compared against;
- who ran it and when.
Approve a version
When a version passes your review, press Set as approved (on the result, or under Scanned models). Every later scan of that model gets one of these statuses:
| Status | Meaning |
|---|---|
| Matches approved | the same commit you approved, and nothing found |
| Changed | a different commit and nothing found: a new version was published; review it and approve it |
| Tampered | the scan found safety removal |
| No approved version | nobody has approved a version of this model yet |
Approving a version that was found tampered needs an explicit confirmation and is written to the audit log.
Reports
Every finished scan has a Report (a printable page) and Signed JSON: the same report signed with our Ed25519
key. Anyone can check that a report is genuine and unedited, online with POST /v1/integrity/verify or offline with
our public key (see Signed reports).
Your organization and team
Everything you scan, approve and download belongs to your active organization (switch it from the account menu). Everyone in the organization sees the same models, scans and reports; other organizations never do. If you are not in an organization, you work in a personal workspace.
Each scan records who started it and whether it came from the web app or the API. Filter to only mine on the
web, or ?mine=1 in the API.
| Role | Can |
|---|---|
| Viewer | read results, history and reports |
| Member | everything a viewer can, plus run scans and approve versions |
| Owner / Admin | everything a member can, plus connect or remove the Hugging Face token |
Roles come from your organization membership. Approvals and source changes appear in the audit log on Connect.
Limits
| Limit | Value |
|---|---|
| Your daily allowance | 10 scans per person per day, web app and API together (Limits and usage) |
| Per organization | 50 scans per 24 hours, up to 10 waiting at once |
| Model size | checkpoints up to 20 GB |
What we store
Measurements, the verdict and confidence, both commit hashes, the method and calibration versions, who scanned and who approved what. Never model weights or activations. Hugging Face tokens are encrypted at rest, used only to read your repositories during a scan, and never shown again.