Show what your model saw, and who was in the loop.
Create a time-stamped record of datasets, prompts, human oversight and training checkpoints without uploading model weights or sensitive training data.
Evidence record
AI Governance
The problem
What makes this hard to prove today
Most evidence problems are not about whether a file exists. They are about whether you can show what it was, when it existed and that it has not changed since.
Training data changes over time
Proving the exact corpus state at training time is hard after the fact, especially when datasets are updated in place.
TDM opt-outs need evidence
Organisations need to show they respected machine-readable reservations in place at the relevant time.
Human oversight must be documented
US and UK authorship rules and the EU AI Act require a record of human involvement, not just an output.
Synthetic media needs authentication
Deepfakes make authentic corporate media and executive communications harder to trust.
What you get
An evidence pack you can actually use
IP Vault produces the sealed record. Your team keeps control of the files and decides who receives the verification link.
A dataset or checkpoint seal
Hash dataset versions, DVC or lakeFS checksums, or model weight files without uploading them.
A prompt and curation record
Capture prompts, model versions, seeds and human review steps in the declaration.
A TDM reservation snapshot
Record the machine-readable signals in place at the training cutoff.
An oversight record
Document who reviewed what, when, and why a version changed.
How it works
Three steps, no new storage platform
Define the record
Choose the dataset snapshot, checkpoint, prompt log or oversight event that needs to be fixed in time.
Seal the snapshot
The digest and declaration are timestamped externally and anchored to Bitcoin.
Share the evidence
Give auditors, counsel or partners a verification record without exposing the underlying model or data.
Regulatory context
Built for AI record-keeping and provenance
IP Vault creates evidentiary records. It does not certify AI Act conformity, audit a model or provide legal advice.
Who it is for
Built for teams that have to show their work
Common questions
The short answers
Do you need our model weights?
Can we prove a TDM opt-out was in place?
Does this make us AI Act compliant?
Can we record human review?
Create the record now, decide who sees it later
Start with a single protection run. Keep the archive, share the verification link, and add revisions when the work changes.
Other solutions