Data Providers

Sell the insight. Protect the individual.

A data provider answers to both the individuals in the data and the customers who buy it: data needs to be properly de-identified, but still useful for the tasks it was bought for.

Two obligations

Privacy upstream, utility downstream

The people your data describes are owed privacy. The customers who buy it are owed a dataset that is faithful enough to use.

It looks like a choice has to be made between privacy and utility, since every step that ensures privacy seems to reduce usefulness.

A semantically twinned record does both. It changes a person's details in a way that preserves fidelity. You can protect the people in your dataset while preserving its value to buyers.

A dataset worth buying behaves like the real thing. A twinned one does.

The twin changes who the record is about, without altering what the customer is buying.

What buyers get today

Safe and thin, or detailed and hollow

Most providers share sensitive data in one of two ways.

  • Summary statistics keep identities private but lose row-level detail.
  • Masked records keep the rows but lose the detail in them: the name, date and postcode go, and with them the patterns a buyer wanted to find.

Masking also leaves a quieter risk behind. It catches the obvious identifiers but misses the indirect ones - contextual facts like a job title or rare condition that, put together, point to one person.

Record · Consumer panel

Field Original stays with you Twin what the model sees Preserved
Panellist Priya Shah Anika Rao gender, name origin
Age 34 35 age band
Postcode LS6 3HJ LS7 2QT region, area type
Household income £42,000 £43,500 same band
Purchases last month 14 14 unchanged
Every field is still complete and plausible. The purchase count is unchanged, so it is identical on both sides: nothing about it identifies anyone.

Privacy

Nobody in the dataset can be picked out

Nakato reads each record for meaning. It finds the indirect identifiers as well as names and numbers, and twins those too. What this gives you is a dataset where every value is plausible and coherent, without pointing to a real person. Assess identification risk before sharing records.

One record, twinned: same shape, same context, different face.

Utility

A truer picture of the data

A twin maintains the age band, format, and relationships between fields. Buyers can run record-level analysis and train models on it, rather than working around blanks. Check whether de-identified data is still useful.

The layer runs within your own environment, on the data you already hold, and works with the architecture you already use. How it works.