Book a demo Start for free

For AI governance

A clear record of the external models your software depends on.

See which Hugging Face models are in use, the revision each one is recorded at, and the evidence for every upstream change.

  • Inventory with file and line
  • Before-and-after evidence
  • No automated verdicts
Model dependencies
4 dependencies 2 pinned 1 floating 1 unknown
ModelRevisionFound in
example-org/chat-7bPinned 3f9a2c1app/models.py:4
example-org/embed-largeFloating mainapp/models.py:5
example-org/vision-encoderPinned 91c0d7eAdded by hand
example-org/rerankerUnknownjobs/rank.py:31
Example

What reviewers can see

  1. 1

    An inventory of model dependencies

    One list of the external models your software uses.

    • Add models by hand, or import them from a project audit
    • Audited dependencies keep the file and line where they were found
    Audit a project
    Add by hand example-org/vision-encoder
    Import from an audit $ pindiff audit .
    One dependency list
    • example-org/chat-7bapp/models.py:4
    • example-org/embed-largeapp/models.py:5
    • example-org/vision-encoderby hand
    Example
  2. 2

    The recorded revision for each one

    Each dependency is classified so reviewers know whether it can move.

    • Pinned: an exact commit
    • Floating: main, a branch or a tag
    • PinDiff shows the recorded revision and does not claim to know what is deployed
    How revisions are tracked
    Pinned revision="3f9a2c1"

    An exact commit. It always points to the same files.

    Floating main · v1.2 · no revision

    Can resolve to a newer commit on the next download.

    ? Unknown revision=REV

    Set at runtime. Record the revision to see your exposure.

    Example
  3. 3

    Licence, access and remote-code changes

    Material upstream changes are recorded and classified.

    • Licence changes show the previous and new licence when both are known
    • Gating, remote-code configuration, weights and model card are tracked
    example-org/vision-encoder · change history
    Licence
    Previousapache-2.0 Newcc-by-nc-4.0
    Access
    PreviousPublic NewGated
    Remote code
    PreviousNot required NewRequired
    Example
  4. 4

    Evidence behind each conclusion

    Every change links to the state PinDiff observed before and after.

    • File hashes for weights, without downloading them
    • Upstream significance shown separately from your exposure
    What evidence is kept
    Evidence · Weights changed
    Before · 3f9a2c1 model.safetensors sha256 a41c…9e02 4.97 GB
    After · 8b7d4e0 model.safetensors sha256 7d3b…11fa 4.97 GB
    Upstream significanceCritical
    Your exposure (pinned 3f9a2c1)No automatic effect
    Example

Evidence for your reviewers, not a verdict

Decisions stay with your team. PinDiff supplies the record they review.

PinDiff does

  • Lists the external models in use
  • Records the revision each one is recorded at
  • Stores before-and-after evidence for upstream changes
  • Shows whether a change can reach the recorded revision

PinDiff does not

  • Label models safe, unsafe, approved or rejected
  • Certify legal or regulatory compliance
  • Run approval workflows or a policy engine
  • Cover private repositories or other model hubs

Not available yet

Start building your model dependency record.

The Free plan covers up to 10 models. For higher limits or a walkthrough, talk to us.

Already have an account? Sign in