Synthetic Person Profile — Variance & Fidelity Report

An independent quality-assurance review of one snapshot of synthetic person-profile records. Tests uniqueness, fidelity to public reference distributions, and internal consistency across 2,000,000 records produced in 10 independent regenerations.

Executive summary

1. Dataset characteristics

10 independent datasets of approximately 200,000 records each. Each dataset is generated with a distinct base seed; downstream sections of this report use the pooled population (n = 2,000,000) for distribution tests, and the 10 individual datasets for the variance analysis in section 4.

Uniqueness with 95% Wilson confidence intervals

Identity-tuple collisions are structurally expected at scale — see the birthday-paradox section below for the analytical derivation. Whole-row and ID-field collisions are the diagnostic uniqueness signals.

2. Distribution fidelity

For each marginal distribution where a published US-adult reference exists, the observed proportions are compared against the reference using a formal statistical test. A MATCH indicates the null hypothesis (observed and reference are drawn from the same distribution) is not rejected at α = 0.05.

Categorical attributes (chi-squared goodness-of-fit)

Numeric attributes (two-sample Kolmogorov–Smirnov)

Observed vs reference — histograms

Bars show the observed empirical density; the red line is the parametric reference density at the same support. Lower KS statistics indicate closer match in distribution shape.

3. Joint structure

Marginal fidelity is necessary but not sufficient; an analyst will also ask whether the joint structure between attributes matches the real world. Two cross-tabs:

3a. Diabetes prevalence by age band

3b. Mean BMI by age band

3c. Inter-attribute Pearson correlations

Practical-equivalence band ±0.15. Sign mismatches (observed and reference of opposite sign with reference |r| ≥ 0.1) are reported separately as a stronger flag.

4. Variance across regenerations

A key quality of a synthetic data product is that regenerating the dataset yields a meaningfully different sample, not a near-copy. Tested by computing the Jaccard overlap of identity tuples across all C(10,2) = 45 dataset pairs.

Per-attribute distribution drift across regenerations

Total-variation distance between each pair of dataset's empirical distributions. Lower values mean more stable marginals across regenerations.

5. Uniqueness & birthday-paradox analysis

The identity-tuple uniqueness percentage decreases at scale because the (given_name × family_name × date_of_birth) value space is finite. The expected collision count under uniform sampling from that space is computed and compared against the observed count.

6. Format validity

7. Quality assurance summary

All cross-field invariant checks pass on this snapshot. 2,000,000 records, 0 violations.

The check battery covers: BMI consistency with height & weight; ZIP code consistency with state; pregnancy_status consistency with sex_at_birth; insulin use consistency with diabetes_status; prescription-count consistency with chronic-medication flags; age within the adult range; and ~30 other invariants.

8. Limitations

• Identity-tuple collisions are structurally expected at scale. At n = 2,000,000 records drawn from a finite name × DOB pool of size ~51,244,200, the expected number of (given_name, family_name, date_of_birth) collisions is approximately 39028.8; the observed count of 46,283 matches the birthday-paradox prediction (AS EXPECTED). This is a property of finite name pools, not a generator defect. Use the ID field for guaranteed uniqueness.
• en-GB and en-IN locales use en-US health reference data as fallback, disclosed via a per-record metadata field. en-US is the only locale with native health-reference calibration in this engine version.
• Cohort-level fidelity is out of scope for this report. Longitudinal events (encounters, prescriptions filled, lab observations over time) are produced by separate engines and reported separately.
• This is engine v0.4 — early in a versioned roadmap. Each release narrows specific gaps; this report is the public quality ledger for one snapshot generated by one engine version. The published version stamp on every record makes future reports trivially comparable.

9. Privacy & re-identification

By construction, every value in this dataset is drawn from a synthetic distribution. No record corresponds to a real person; no field is derived from any real record. As a consequence:

• The dataset is HIPAA-equivalent on its face: there are no protected health identifiers because there are no real-world identifiers at all. (Formal HIPAA Safe Harbor attestation by a statistician is a separate compliance artifact; this report makes the structural argument.)
• Re-identification via attribute combination (the "linkage attack" threat against anonymized real data) does not apply. There is no underlying real distribution whose tails could uniquely identify an individual.
• Identity-tuple collisions (multiple records sharing first_name + family_name + date_of_birth) are an expected statistical artifact of drawing from finite name pools, not a privacy concern. See the birthday-paradox derivation above.
• Sensitive-format fields (ssn_last_four, employer_ein) use structurally invalid prefixes that cannot collide with real-world assignments, so a test ID never accidentally resembles a real one.

10. Reproducibility

This dataset can be regenerated via the public API:

POST /v1/mock/person
{
  "count":   200000,
  "seed":    <any uint64>,
  "locale":  "en-US"
}

Each request is byte-identical for the same seed at the engine version listed in the header card.