Assurance and compliance

What the security review will ask. Already answered.

Sign-off waits on a familiar set of questions. Here they are, answered by how the layer is built rather than by promise — so the person who owns the data can put their name to it, and the person who owns the tooling already has the answer.

Where does our data go?

Your real data never leaves your environment.

Transformation runs inside your VPC or on-premise, on a small language model, with no dependency on a big API provider. Nothing about your existing data classification has to change.

What happens when detection is unsure?

Fail-closed by design.

Entities below the confidence gate are held back, never passed through. The tool is designed so the model has no reasonable means of identifying your customers.

Can we evidence what was substituted?

Every substitution is auditable.

A full audit trail records every transformation, ready for supervisory review.

Which models can we use?

Works with any model.

The layer sits in front of the model, not inside it. Switch providers without rewriting your governance.

Works in front of

  • Claude
  • GPT
  • Gemini
  • Mistral
  • Fine-tuned models

What does detection actually catch?

Context, not keywords.

50+ entity types, including indirectly identifying information that pattern matching misses. Meaning is read, not matched.

Compliance

Evidence, not assurances.

Every benchmark is published and checkable. Every substitution is written to the audit trail. And nothing about your existing data classification has to change.

EU AI Act
Transparency evidence for high-risk systems, produced as the layer runs.
DORA
Operational resilience inside your own environment, not a new external dependency.
FCA / PRA
Supervisory evidence on demand — every substitution and its reversal on the record.
GDPR
Data protection by design, with pseudonymisation as the regulators define it.
SOC 2
Framework-aligned controls, documented and checkable in the audit trail.

Built against these from the start. Not retrofitted.

The layer is designed so the model has no reasonable means of identifying your customers.

Put your real data to work.

Your real data never leaves your environment. The model only ever sees the twin, and you still get the accuracy you needed.

© 2026 Nakato. All rights reserved.Equivalent. Not identical.