OIS Research Services

Research on sensitive data that never leaves the office

oissyntheticdata 3.0.0: an open-source tool, in Python and in R, for government offices and the researchers who work with them

Dr Yohanan Ouaknine

scroll, or press ↓

01 / the problem

The sensitive files cannot leave the office

An example: a ministry keeps its service requests in a folder of Excel and CSV files, with ID numbers, names and addresses. The question is simple: how many days does it take to close a request, in each district?

Hand the files to an outside researcher?

Not possible. They hold personal data.

Write the analysis without seeing the files?

Nobody can write code that runs without knowing the column names, what the columns hold and how the tables link to each other.

The example is inventedoissyntheticdata 3.0.0
02 / the idea

Send a description of the data, not the data

Inside the office, one script that needs only the Python standard library reads the files and writes a small description, as JSON files: column names, types, distributions and the links between tables. It contains no row of the data.

A row in the file, inside the office

ID number
012345678
Name
Israel Israeli
District
North
Opened
14 March 2026

The same columns, in the description that leaves

ID number
Digits, length 9
Name
Letters; the length only
District
North 412 · Centre 1,038 · South 367
Opened
March 2026: 163 requests
The example is inventedoissyntheticdata 3.0.0
03 / the flow

The flow, in four steps

1

On the premises: the profile

Run one file in the data folder. It writes an aggregate description of the files.

2

The office sends the folder

Only the to_send folder goes out, by ordinary e-mail. It contains no row of the data.

3

At home: synthetic files and your script

The same structure, and every value invented. You write your analysis script on these files.

4

Back on the premises: the same script, on the real files

The script does not change; only now are the numbers findings.

Steps 1, 2 and 4 happen on the premises; only step 3 happens outsideoissyntheticdata 3.0.0
04 / on the premises · steps 1 and 2

On the premises: one file, one command

Copy gov_one.py into the data folder and run:

python gov_one.py

The tool reads every Excel and CSV file, works out how the tables link to each other, and writes a folder called to_send. Only that folder goes out, by ordinary e-mail.

What is in the to_send folder

profile_<table>.json
One file per table: counts, shares and rounded quantiles only
schema.json
The links between the tables
profile_summary.md
Exactly what leaves, column by column. The person who approves the transfer reads it first.
Nothing to install and no internet on the premisesoissyntheticdata 3.0.0
05 / at home · step 3

At home: synthetic files, and your script on them

First: synthetic files

The same tool, the same command, next to the to_send folder:

python gov_one.py

The same file names, sheets, headers and row counts, the same links between tables, and similar distributions. Every value is invented.

Then: the analysis script

Write your analysis in the script template and run:

python analysis_template.py .

The built-in toolbox: statistical tests, regression, interrupted time series, and an HTML report that opens offline.

The report is stamped “Synthetic test data — not findings”oissyntheticdata 3.0.0
06 / back on the premises · step 4

The same script, on the real files

Bring the tool and your script, put them next to the real files, and run. The script does not change; only now are the numbers findings.

At homeSynthetic filesYour scriptA check that the code runs
On the premisesThe real filesThe same scriptFindings
With the real files in the folder, the script reads them and the report loses its synthetic stampoissyntheticdata 3.0.0
07 / what leaves

What leaves the office, and what does not

In the data
In the JSON files that leave
In the dataID numbers, names, phone numbers, addresses, free text
In the JSON files that leaveShape only: the length, and whether it is letters or digits
In the dataCategories (district, status, request type)
In the JSON files that leaveLabels and counts; a label seen fewer than five times loses its name
In the dataNumbers
In the JSON files that leaveNine rounded quantiles, the mean and the spread
In the dataDates
In the JSON files that leaveCounts per month and per day of the week; never the date itself
In the dataLinked tables
In the JSON files that leaveHow many rows each key has, and what share of the parent records have any
No row is copied. No count below 5 leaves, and that minimum can be raised.oissyntheticdata 3.0.0
08 / what for

Synthetic files: for testing the code, not for findings

What they are good for

  • Writing the code against files with the same structure as the real ones
  • Checking that the script runs properly, from reading the files to the report
  • Fixing errors at home, before going to the office

What they are not good for

  • Findings. A result computed on them is never a finding.
  • Every reported number comes from running the script on the real data, on the premises.
Every value in the synthetic files is inventedoissyntheticdata 3.0.0
09 / install

Install, in Python or in R

Python

pip install oissyntheticdata
  • Version 3.0.0 on PyPI, released 12 November 2026
  • Python 3.8 or later, standard library only
  • MIT licence

R

install.packages("remotes")
remotes::install_github("yohananouaknine/oissyntheticdata-r")
  • Version 3.0.0 on GitHub
  • Base R only, R 4.2 or later
  • The same JSON files as the Python package
  • MIT licence
Nothing is installed on the premises. You carry in two files only: the tool and the script template.oissyntheticdata 3.0.0

Who is behind the tool

Dr Yohanan Ouaknine

Information scientist (PhD, Bar-Ilan University), formerly Head of the Research Branch of the Israel Prison Service. He lectures on research methods, is a research fellow at the Open University of Israel, and is the founder of OIS Research Services.

Arrange a call