Validation
Synthetic clinical data should be measurable.
Synset reports validation by tier, separating certified trajectory support, accepted-only audit results, withheld-prior validation, self-consistency testing, and held-out future validation when available.
Validation Tiers
Tier 1: Held-Out Future Validation
Uses real observed future data intentionally hidden during generation.
- Acute: pending local row-level future trajectory table.
- Chronic: pending sufficient row-level repeated-visit future data.
Tier 2: Independent / Withheld Validation
Uses independent validators or withheld empirical-prior constraints not used as the final public claim of patient-future accuracy.
- Acute: 0 / 14,143 accepted-only validator violations.
- Chronic: 0 / 3,953 final withheld-prior violations after release gating.
Tier 3: Certified Self-Consistency
Masks later certified synthetic segments and verifies regenerated continuations remain within the certified support envelope.
- Measured for acute and chronic.
- Internal stability measure.
- Not ground-truth accuracy.
Acute results
Support-certified horizon: 96 sequential hourly transitions.
- Support-certified horizon
- 96 sequential hourly transitions.
- Accepted-only validator audit
- 0 / 14,143 accepted acute candidates.
- Conservative upper bound
- ~0.021%.
- Yield improvement
- ~7x final-horizon certified yield with reanchoring.
Required caveat: These are not held-out future patient accuracy claims.
Chronic results
Current stage: aggregate-prior-distilled prototype.
- Current stage
- aggregate-prior-distilled prototype.
- Withheld-prior validation after optimization
- 0 / 3,953 final audit checks.
- Conservative upper bound
- ~0.076%.
Required caveat: This is withheld empirical-prior validation, not held-out patient-future accuracy.
Build validation into the generation workflow.
Candidate trajectories are generated, checked, filtered, and audited before they are treated as releasable synthetic data.