Validation
Validation that distinguishes rows, trajectories, and future claims.
Synset reports validation by tier. We separate what has been directly measured from what remains pending, and we avoid treating internal plausibility metrics as real-world future accuracy.
Validation tiers
Tier 1
Held-Out Future Validation
Compares accepted synthetic continuations against real future observations that were hidden during generation.
- Acute: pending row-level observed future trajectory table.
- Chronic: pending sufficient row-level future data.
Tier 2
Independent / Withheld Validation
Tests accepted outputs against validators or empirical-prior constraints not used as held-out future patient outcomes.
- Acute: 0 / 14,143 accepted-only validator violations.
- Chronic: 0 / 3,953 final withheld-prior violations after release gating.
Tier 3
Certified Self-Consistency
Checks whether regenerated continuations remain consistent with certified trajectory envelopes.
- Measured for acute and chronic.
- Internal stability metric.
- Not ground-truth accuracy.
| Regime | Tier 1 observed future | Tier 2 withheld validation | Tier 3 self-consistency |
|---|---|---|---|
| Acute | Pending | 0 / 14,143 | 0 escapes |
| Chronic | Pending | 0 / 3,953 | 0 escapes |
Acute trajectory validation
Acute-care trajectories over four days of hourly state changes.
Synset can generate acute-care trajectories that remain plausible over 96 hourly transitions. In post-generation validator audit, no violations were observed among 14,143 accepted acute candidates.
- Support-certified horizon
- 96 hourly transitions
- Accepted-only audit
- 0 / 14,143 violations
- Upper confidence bound
- ~0.021%
- Reanchored yield improvement
- ~7x
Required caveat: this is not yet a held-out future patient-outcome accuracy estimate.
Chronic/outpatient prototype validation
Withheld empirical-prior validation, not future prediction.
The chronic prototype is evaluated against empirical longitudinal-prior constraints that were withheld from the final release gate. This measures adherence to withheld prior structure, not patient-future prediction accuracy.
- Validation type
- withheld empirical-prior validation
- Final violations
- 0 / 3,953
- Upper confidence bound
- ~0.076%
- Status
- not held-out patient-future validation
Required caveat: chronic validation is not held-out patient-future validation.
Claim boundaries are part of the product.
Synset distinguishes between accepted candidates, released trajectories, withheld validation checks, self-consistency, and true held-out future validation. The system should make those boundaries visible instead of flattening them into one score.
Plain-language summary
Build validation into the generation workflow.
Candidate trajectories are generated, challenged, filtered, and audited before they are treated as releasable synthetic data.