Models come from open repositories, vendors and your own training runs. We trace each one from source to production, then help you add checks that stop unapproved versions shipping.
- Base modelOpen weights from an approved mirrorHash recorded
- Training dataSupport tickets dataset, version 7Reviewed
- Fine-tuneRun 118, job identity ml-trainLogged
- EvaluationEval set 12, all checks passedTied to file
- SignatureML platform team, 2 approversSigned
- Live endpointsha256 9f2c…a71eMatches
Follow one release
Six places a model changes hands.
Pick a step to see what we check there, and what usually goes wrong.
What we check
Where the weights came from, their licence and a fingerprint recorded on arrival.
What goes wrong
A look-alike repository serves a modified copy.
The record we look for
source: approved mirror · sha256 recorded
Some model files can run code the moment they load.
Older file formats can carry instructions as well as weights. A safer format stores only the numbers.
- $
python load_model.py support-summariser.pkl - >
pickle files can run code while they load - !
found: a call to os.system inside the file - ×
load refused: untrusted pickle - $
convert --to safetensors support-summariser.pkl - ✓
loaded 4.2 GB of weights, no code run
Approval only counts if the same file ships.
Model files look alike from the outside. A fingerprint check at deploy tells them apart.
sha256 9f2c…a71eSignedsha256 41bd…07c2UnsignedWhat to do first.
- 01
Record a fingerprint for every model.
Write down where each model came from, and its hash, the day it arrives.
- 02
Sign a model when you approve it.
Approval should belong to one exact file, not to a name or a version tag.
- 03
Check the signature at deploy.
Production only runs files that match a signed approval.
- 04
Keep the last good version ready.
If something looks wrong, roll back to the last approved file.
How we help.
Integrity checks prove a model wasn't swapped. They don't prove it behaves well; that needs its own evaluation, and we'll tell you when you need it.
We trace one release.
From source weights to the live endpoint, including who can change each step.
We help you add the checks.
Signing, deploy checks and rollback, built into the pipeline you already use.
We prove they work.
We try to ship an unapproved model, and show you where it gets stopped.
We keep watching.
With one of our security programmes, we watch your release pipeline for changes that need a fresh look.
Model supply chain
Trace your next model release.
We'll follow it from source to serving and show where trust is assumed instead of checked.