Watch a semantic twin at work

Your team wants to use an external tool. The data it needs is too sensitive to share. Mask the name and email, and a job title can still give someone away. Redact too much, and you lose the facts the tool needs to get the answer right.

Nakato detects sensitive details inside your company. Take Sarah's request. The team promises a reply within two days. She's waited three. At runtime, it reads the task and decides which details to replace and which facts to keep. Sarah's name and email identify her. An indirect clue, like being the only manager in the York office, can point to her too. Sarah Jones becomes Emma Reed, with a matching email. The role note becomes manager in the Leeds office. The 72-hour wait stays with the request.

That’s a semantic twin. It’s contextually identical for this task, and Sarah is unidentifiable to the tool. Only the twin leaves your company. With the waiting time intact, the tool can flag the overdue request.

The answer returns. Nakato restores the real details locally. On a public reasoning benchmark, redaction, synthetic replacement and tokenisation kept at most 63% of the raw model’s score. Semantic twins kept 96%. That gap can be the difference between second-guessing a tool’s answer and trusting it enough to act. Wherever you work with sensitive data, Nakato protects identities and preserves the context your tools need for accurate answers. Put your data to work. The real details stay with you.