Quality certificate¶
The Quality certificate is a one-screen verdict on a finished project: is the synthetic data fit to deliver, how close is it to your real data, and what is it safe to use it for? Read it before you send anything to a client.
It is not the same thing as the Quality and performance report, which is the Excel audit you download. The certificate is the summary you read on screen; the report is the detailed file behind the numbers.
Where to find it¶
Open a finished project (see Monitoring & results) and look for the Quality certificate panel on the project detail screen. Click its header to open or close it.
Every project that finishes now gets its certificate automatically — there is nothing to switch on. Automatic mode projects get one too.
Not in production yet
These two changes are built but not live on app.ninfinite.studio yet, so what you see today may differ slightly:
- Collapsed by default. The panel will start closed so the project detail stays short; click the header to open it.
- Older projects. Projects that finished before 8 October 2026 may not have a certificate. A Compute quality certificate button will create one on demand. While it works the panel says it is computing; if it fails, the panel shows a reference code (the one support needs, see Waiting, queues and errors) and a Retry button.
The verdict¶
Not in production yet
The verdict, and the Real base wording in the next section, are built but not live on app.ninfinite.studio yet. Until they ship, the panel shows a single badge called Viable, which only reflects the amount of real data (see Real base), not the quality of the synthetic data.
At the top of the panel one verdict says whether the data is fit to deliver. It combines three things: the real base, the quality scores and the delivery checks. Under it the panel lists the reasons, most important first.
| Verdict | When you see it | What to do |
|---|---|---|
| Ready to use | All of these: the real base is Sufficient, every delivery check passed, and the Core score is 80 or more. | Deliver. The line Recommended for names the kinds of analysis whose score is 80 or more. |
| Use with caution | Anything in between. The reasons can be: fewer than 300 real complete cases (or fewer than 150); more than 50% synthetic rows; Core score from 50 to 79; a small synthetic part (under 100 rows, so scores vary between runs); or the Core score could not be calculated. | Read the reasons and deliver only for the uses listed under Recommended for. Below 150 real cases that means totals only, not breakdowns by segment. |
| Not recommended | Any of these: more than 70% of the rows are synthetic, at least one delivery check failed, or the Core score is below 50. | Do not deliver as it is. For a failed check, start with Post-processing. For a thin real base, add real respondents. For a low Core score, look at the five scores to see which kind of analysis is weak. |
Real base¶
Under the verdict, one line shows how many rows are real and how many are synthetic, for example 755 real · 252 synthetic (25% synthetic). A badge titled Real base rates that mix.
The Real base badge only measures how much real data is behind the delivery. It says nothing about the quality of the synthetic rows; that is what the scores below are for.
| Badge | When you see it | What it means |
|---|---|---|
| Green Sufficient | At least 300 real complete cases (respondents who finished the questionnaire) and no more than 50% synthetic rows | Enough real data to build on, and the synthetic part is a modest addition. |
| Yellow Limited | Fewer than 300 real complete cases, or more than 50% synthetic rows | Usable, but read the notes. Below 150 real cases the certificate recommends using the data for totals only, not for breakdowns by segment. |
| Red Too thin | More than 70% of the delivered rows are synthetic | Too little real data behind the result to stand on its own. |
How quality is measured¶
The question is: how close is the synthetic data to the real data? A raw difference means little on its own, because even two real samples of the same size never match exactly. So the certificate first measures that natural variation:
- It splits your real data into two random halves, many times, at the sizes of your project, and records how much two real samples differ.
- It measures the same difference between the synthetic data and the real
data, in four families of metrics:
- Percentages and spread of each question.
- Relationships between questions, plus a test of whether a classifier can tell real rows from synthetic ones.
- Applied analyses: rankings and cross-tabs by profile.
- Diversity: rare answers are kept and respondents are not copied.
- It places the synthetic difference on a 0 to 100 scale against the real one.
What the numbers mean. 100 means the synthetic data differs from the real data no more than a typical second real sample would (the median of the real-versus-real comparisons). 0 means it differs more than 95% of those comparisons. In between the scale is linear; only diversity uses a smooth curve, where 50 sits at the 95th percentile.
Example. A Topline score of 85 means the question percentages differ from the real data about as much as two real samples of the same size would: a little more than the typical pair, well within the normal range of pairs.
The five scores¶
Below the verdict you get one Core score from 0 to 100, followed by five scores, one for each kind of analysis you are likely to run on the data. Each task score is a weighted mix of the four families above. Core is the mean of the five. Each of the five is coloured by band:
| Colour | Score |
|---|---|
| Green | 80 or more |
| Yellow | 65 to 79 |
| Orange | 50 to 64 |
| Red | Below 50 |
| Score | What it tells you | Mostly built from |
|---|---|---|
| Topline | How close each question's percentages are to the real data. Good for totals and headline numbers. | Percentages and spread |
| Profiles / cross-tabs | How well the data holds up when you cut it by profile groups such as gender, age or country. Good for breakdowns. | Applied analyses |
| Multi-response rankings and funnels | How faithfully multi-select answers, rankings and brand funnels (aware, considered, bought…) are reproduced. Good for brand tracking. | Applied analyses |
| Relationships and drivers | Whether the links between questions that you would use for driver analysis are preserved. Good for driver analysis. | Relationships |
| Segmentation | Whether the same segments you would find in the real data show up in the delivered data. Good for segmentation work. | Relationships and diversity |
A kind of analysis is listed under Recommended for when its score is 80 or more. If the real base is below 150 cases, only Topline can be listed.
Delivery checks¶
A list of pass/fail checks on the delivered rows. Each one shows Passed or Failed, and a failed check makes the verdict Not recommended:
- Structural rule violations — numeric relationships found in your file (for example, parts that must add up to a total) are respected.
- Empty questions — no question is left completely blank in the synthetic rows.
- Exact duplicate rows — no two synthetic rows are identical.
- Rows copied from a real respondent — no synthetic row is a copy of a real person.
- Answers within the study's options — every generated answer is a code that exists in your real data for that question, so no invented answer options appear.
If a check fails, Post-processing is the first place to look: logical consistency repair and Remove duplicates act on exactly these problems, and you can download again without regenerating anything.
Limits¶
- An internal estimate. The comparison is against the same real data the engine learned from, not against respondents it never saw. That makes the scores somewhat optimistic. External validation with held-out data is available on request.
- Few synthetic rows. With a small synthetic part (under 100 rows) there is little data to measure, so the scores move a lot from one run to another. Treat them as a rough guide, not a precise figure; the verdict flags this as a reason for caution.
Notes and recommended uses¶
Notes point out anything that limits how far you can trust or use the result — for example that the synthetic part is small, or that the file has no numeric questions to check. Recommended uses says what the data is good for given everything above, for example totals only.
Same inputs, same file¶
The generation is reproducible: the same project, on the same data, with the same random seed produces the same file. The seed is the Random seed in Compare executions and in the Statistical Refinement settings of Post-processing.
Automatic mode deliberately uses a different seed for each of its five versions, then delivers the one with the best internal quality score.