The Data Centrifuge is aytm's automated data quality engine. It reviews every respondent in your survey against a set of fraud and inattention checks, rates how suspicious each one looks, and sets the worst offenders aside so they never reach your results.
In This Article
- Terminology
- What the Data Centrifuge Is
- What It Looks For
- Suspicion Levels and Quarantine
- Reviewing Flagged Responses
- Common Questions
Terminology
| Term | Definition |
|---|---|
| Data Quality Report | The panel on your Stats page that brings together everything aytm knows about the quality of your sample. The Data Centrifuge is one section inside it. |
| Vector | A single detection signal—one specific kind of low-quality answering behavior the engine looks for. |
| Suspicion level | How strongly the evidence points to a bad response, expressed as High, Moderate, Low, or Minor. |
| Quarantine | Setting a flagged respondent aside so they're excluded from your analyzed data. Quarantined responses aren't deleted. |
| Release | Putting a quarantined respondent back into your analyzed data. |
What the Data Centrifuge Is
Every completed response is reviewed independently against each active check. Those results are combined into a single suspicion rating for that respondent, and the respondents with the strongest evidence against them are pulled out of your analyzed dataset. You can review the evidence behind every flag, put back anyone you disagree with, and reject anyone the engine didn't catch.
To find it, open your survey's Stats page, switch to the Quality tab, and scroll to the Data Centrifuge section of the Data Quality Report.
When It Runs
The Data Centrifuge works in two passes. While your survey is still collecting, it reviews responses as they arrive, so you get a live read on quality mid-field. Once fielding closes, it runs again across your full respondent pool—this is the authoritative pass, and it's what your final quality numbers reflect.
That second pass matters because a number of the checks judge a respondent against everyone else in your sample rather than against a fixed rule. Those comparisons can't be made until the whole sample is in, so they'll show as not applicable until the final run.
Where It Sits in the Data Quality Report
The Data Centrifuge is one of five sections in the Data Quality Report, and it's worth knowing what the others cover:
- Panel—how many responses you collected.
- Traffic Sentinel—suspicious and duplicate entry attempts that were blocked before they ever entered your survey, plus pre-qualification drop-offs.
- Data Centrifuge—quality review and quarantine of the respondents who did get in.
- Human Review—manual rejections, open-end rejections, and reconciliation.
- Performance—completion rate, drop-offs, and the respondent experience rating your survey earned.
Think of Traffic Sentinel as the front door and the Data Centrifuge as the back door. One keeps bad traffic out; the other catches what made it through.
What It Looks For
Checks are organized into categories. Each category holds one or more individual checks, and each one targets a different kind of low-quality or fraudulent answering behavior.
| Category | What it detects |
|---|---|
| Twinned Responses | Respondents who duplicated each other—the same closed answers, the same open-end text, or the same uploaded image. |
| Inconsistent Responses | Answer combinations that almost never occur together in this audience. Each answer alone may be normal; together they form an improbable profile. |
| Random Responses | Answering that looks like random clicking rather than genuine preference. |
| Pacing Anomalies | Respondents who moved far faster—or far slower—than the survey realistically allows for. |
| Response Patterns | Straight-lining—the same answer in the same position on every screen, with no underlying logic. |
| Interrupted Sessions | Sessions with a long unexplained gap, suggesting the respondent stepped away and came back. |
| HB Model Fit Issues | MaxDiff answers that look chaotic rather than preference-driven. Applies only to surveys containing a MaxDiff question. |
| Attention Checks | Respondents who failed a quality-control question you planted—for example, one that instructs them to select a specific answer. |
| Verbatim Anomalies | The largest category. Checks on the quality of open-ended text, including nonsense, off-topic answers, repeated answers, offensive language, and text that appears to have been written by an AI generator rather than a person. |
| Digital Body Language Signals | Signals from how a respondent interacted with the page rather than what they answered—for example, pasting text into an open-end box instead of typing it. |
⚠️ Note: Some checks depend on your survey containing the right ingredients. Attention Checks only have something to find if you included an attention-check question, and HB Model Fit Issues only apply if your survey contains a MaxDiff question. When a check can't run, it shows a dash rather than a zero—see Common Questions.
Suspicion Levels and Quarantine
Each flagged respondent is assigned a suspicion level based on how much evidence stacked up against them:
- High—multiple strong signals, or one severe one. Almost always genuine bad data.
- Moderate—enough converging evidence to remove by default.
- Low—some evidence. Kept in your data, but worth a look if your topic is sensitive.
- Minor—a single light signal. Kept in your data.
Respondents are quarantined automatically when auto-quarantine is turned on for your survey. When it's off, the engine still reviews and reports everything—it simply makes no removals on its own, and you decide what to remove.
Being flagged and being quarantined aren't the same thing. Flagged means at least one check fired. Quarantined means the evidence was strong enough to set the respondent aside.
Quarantined responses are excluded from your analyzed dataset—they won't appear in your charts, crosstabs, or clean-data exports—but they aren't deleted. They stay reviewable, exportable, and restorable.
Common Questions
Why does a check show a dash instead of a number?
A dash means the check couldn't run for your survey—usually because a prerequisite was missing, such as no open-ended questions, no MaxDiff question, or not enough respondents for a statistical comparison. A dash isn't the same as a zero. Zero means the check ran and found nothing wrong.
Why did the same respondent get flagged by several different checks?
Several checks look at related behavior from different angles, so one genuinely poor respondent will often look unusual on more than one dimension at once. That's expected, and it doesn't compound endlessly—suspicion is capped.
Why wasn't anyone flagged for speeding when I know I had speeders?
Some of the pacing checks compare respondents against the rest of your sample rather than against a fixed limit, so they only run once fielding is complete. And because they're relative, they need a normally-paced majority to compare against. If your entire sample moved quickly, nobody stands out.
Can I turn a check off?
You can uncheck a check in the sidebar to preview your data without it. Changing which checks actually count toward removal is a per-survey configuration your aytm contact can adjust for you.
Does my N drop after quarantine?
Yes. Quarantined responses leave your analyzed base, so your reported N reflects the cleaned data.