Erasure evidence, with a receipt.
Most "AI compliance" stops at access logs and prompt-level guardrails. Divinci goes inside the model. Every entity association is a queryable feature; every edit is a portable, auditable patch — and the vIndex records what the model answered before and after, in a form a third party can check.
Research preview. Divinci's model-erasure product is in active research and development and is not generally available. Nothing on this page is a claim that it makes you compliant with GDPR, the EU AI Act or any other law. We publish every test result, including the failures.
Can an AI forget? An honest update
Three minutes on where our research into removing facts from open-weights models stands: what works, what an edit costs, what failed, and what we are trying next.
The worked example in full: Deleting Paris from a Language Model
What the regulation requires
The same artifact answers transparency, oversight, documentation, and erasure requirements.
| Regulation | Requirement | Competitors | vIndex |
|---|---|---|---|
 EU AI Act Annex IV | Technical documentation | Access logs + prompt guardrails only | Draft Annex IV sections from release records and the model-edit log — every gap stated; a draft for your assessment, not a declaration of conformity |
| GDPR Art. 17 | Right-to-erasure with verifiable proof | Delete the source data only | In research: removal by rebuilding the fine-tune without the person, with a signed before-and-after receipt — not yet validated |
| HIPAA §164.312 | Audit trails for protected health information | Manual audit logging | Release records and the model-edit log as audit evidence — not a HIPAA control on its own |
 NIST AI RMF MAP/MEASURE | Risk management framework mapping | Static controls checklist | Continuous, model-native evidence |
 SOC 2 Trust Services | Security, availability, confidentiality | Point-in-time audit | Deployment telemetry as evidence input — Divinci is not SOC 2 certified |
 ISO/IEC 42001 AI Management | AI management system standard | Document-only conformity | vIndex as the system of record — certification not yet held |
What a "receipt" actually looks like
A vIndex patch is a portable JSON file with a SHA-256 checksum. Apply it to change what the model answers; remove it to restore the model bit-for-bit. The patch is the audit log — operators, regulators, and downstream consumers can verify the same operation independently. It records what the model answered before and after; it does not by itself prove the model has forgotten.
Below: an illustration of the format — the fields a Gate-3 DELETE patch carries, and the checksum that makes it verifiable. It is an example of the shape of a receipt, not a measurement we are publishing here.
{
"name": "gdpr-art17-paris-capital",
"version": 1,
"base_model": "google/gemma-4-E2B-it",
"created_at": "2026-04-22T22:34:00Z",
"operations": [{
"op": "delete",
"entity": "Paris",
"relation": "capital",
"target": "서울",
"weight": -1.0,
"layer": 27,
"feature": 11179
}]
}Become a design partner.
We are looking for three design partners who run open-weight models and expect erasure requests. Together we pick one fact, or one person's records, run the removal and the probes against bars set in advance, and share the signed before-and-after receipt. This is research, not a service: the method is still being validated, and you see every result, including failures.
Three limits worth stating up front. Weight-level editing requires an open-weight served model — GPT, Claude and Gemini cannot be edited this way, only retrieved-context can. The receipt certifies what the model answered before and after and the exact patch that undoes it; it does not by itself prove the model has forgotten. On Gemma 4 E2B our own measurements show that removing an association from the index can leave the chat answer unchanged, while an output-level override changes the answer and leaves the old one a rank below — the receipt says which kind of edit you got. And the dossier is a draft for your assessment, not a declaration of conformity: that is a judgement only you can make.
Gemma
Qwen
Llama
Mistral
Microsoft
Twelve published vIndexes — Browse them on Hugging Face
Evidence you can check. A research preview on open-weight models.