I asked 300 synthetic nurses one question. They gave me five answers.

Personas
300
Specialties
ICU · Med-Surg · Home Health
Anchor
HRSA federal workforce data
Collapse, run 1 → 3
82% → 66%

The setup

I built 300 synthetic nurse personas across three specialties, anchored to real federal workforce data (the HRSA National Sample Survey of Registered Nurses), and ran them through an automated validation engine. The goal: could synthetic respondents reproduce the actual reasons nurses leave their jobs?

What worked

On the big relationship, it did great. When a nurse pointed to patient load as their breaking point, they almost always turned out to be a detractor. That is the kind of directional signal synthetic data is genuinely good at, and it matched what the federal data says about why nurses quit.

Where it broke

Then I looked at the eNPS question, the classic 0-to-10 "would you recommend this job." Across 300 personas, the first run produced four unique numbers, and 82% of the nurses piled onto just two of them.

Real nurses don't answer that way. Ask 300 of them and you get a spread: bitter 2s, tired but loyal 7s, the occasional 9. My AI quietly pulled everyone toward a few safe values. The technical term is variance collapse. The plain-English version is that the model mostly agreed with itself.

Same survey, better grounding: the spread returning
eNPS scores (0 to 10) across three runs. Only the sourcing changed. Copper marks each run's most common answer.
Score 2: 11 personas Score 3: 37 personas Score 5: 88 personas Score 8: 135 personas 11 37 88 135 012345678910
Run 1 · baseline sourcing · 4 scores · 82% on two
Score 2: 19 personas Score 3: 46 personas Score 5: 100 personas Score 8: 1 persona Score 9: 134 personas 19 46 100 1 134 012345678910
Run 2 · more sources · 5 scores · 78% on two
Score 2: 37 personas Score 3: 28 personas Score 5: 37 personas Score 6: 63 personas Score 9: 135 personas 37 28 37 63 135 012345678910
Run 3 · triangulated · 5 scores · 66% on two
Run 1 scored n=271 of 300 personas; runs 2 and 3 scored all 300. Percentages are the share of scored responses sitting on that run's two most common values. Hover any bar for its exact count.

The fix, and the honest catch

I kept the survey exactly the same and changed how I grounded the personas. Instead of leaning on a single source, I moved to a triangulated approach, pulling from several independent anchors so the personas weren't all shaped by the same reference point. Then I ran it again. And again.

Each pass, the spread widened. Four unique scores became five. The concentration on the top two answers dropped from 82% to 78% to 66%. The variance was coming back.

But here is the part that matters: even after all that, the scores stayed clustered. Better, clearly. Realistic? Not yet. The model still wanted to bunch people up, and no amount of grounding fully talked it out of that.

What I disclosed

Every limitation in this study is documented in the full report, including:

  • eNPS variance stayed compressed even in the best run. Improved is not fixed.
  • One specialty's trigger question leaned on a proxy source and over-concentrated as a result.
  • Validation against real nurses is the planned next phase and has not run yet.

The lesson

Synthetic data doesn't fail loudly. It never throws an error. It hands you clean, reasonable-looking numbers that lean in a predictable direction: toward the average, toward agreeable, toward the center. If no one is specifically checking for it, it walks straight into a stakeholder meeting.

That is exactly why this work needs a second, independent set of eyes. The vendor who sold you the boost has little reason to mention that the variance is still collapsed under the hood. I don't sell the data, so I'm happy to be the one who checks.

Wondering if your synthetic sample does this too?

One dataset, one week, one plain-language memo on where it holds and where it doesn't. No retainer.

Run the Sanity Check