Prove AI model integrity across version changes, retroactively.
When you update an AI model (v3.5 → v4.0, etc.), you often need to verify later whether past decisions are reproducible or whether the results change. Lemma commits each point-in-time modelId and policyHash at decision time, so past decision logic stays cryptographically traceable even after the model updates.
Three voices from the front line.
- AI engineering / operations
“After a model update, we end up needing to reproduce and verify past decisions retroactively”
- Audit / regulatory
“We want to prove we ran "that policy" on "that model" at the time”
- Legal
“We want to clarify accountability for past AI decisions after the fact”
Hand over the source, or just the facts?
Nothing changes on the floor. Everything changes for the receiver.
① Your team just saves, as always.
- The usual step
- Fill in the record and save
- On save
- A proof is attached (API, behind the scenes)
- The document itself
- never sent
② They just open a link.
- Proven fact
- the same decision before and after a model change (or the diff)
- the model internals, parameters and training data
- not shown
- Login / keys
- not needed
- The document stays private — the record itself is never sent or disclosed.
- Independent verification — the receiver just opens a link. No account, no keys.
- Edits are detected — even a one-character edit fails verification.
At the moment of decision, modelId@timestamp and policyHash are pinned and committed together with the input/output commitments and satisfiesPolicy. This record stays immutable even after the model updates, so past decisions can be verified retroactively against "that model, that policy." Without disclosing model internals or parameters, the integrity of a decision can be independently shown.
See the technical details ↗Why the usual methods fall short.
Only work that needs all three at once — pass without exposing, independent verification, tamper-evidence — is Lemma's domain.
| Method | Pass without exposing | Independent verification | Tamper-evident | What happens |
|---|---|---|---|---|
| Access control / permissions | △ | ✗ | ✗ | “Someone inside could have edited it” remains possible |
| Masking / redacted copies | △ | ✗ | ✗ | Redaction work grows; the original is still unproven |
| Encrypt and store / send | ✓ | ✗ | ✗ | To verify, the receiver needs it disclosed after all |
| Lemma (ZK proof)the only one with all 3 | ✓ | ✓ | ✓ | The receiver just opens a link |
How it works — and how to start.
We help design disclosure scope and retention, run the PoC, and support production.
Start with a 30-minute call.
Tell us one AI system worried about explaining past decisions after a model update, in the first 30 minutes. No disclosure of model internals required.
Find the case that matches your industry and problem
Browse all 36 use cases →TRY LEMMA
Run it yourself.
No sales call needed — start hands-on with Lemma's products.