Evidence

Evidence for release-gated synthetic clinical scenarios.

Synset separates generation from release. Candidate records, trajectories, and text sidecars are checked against explicit admissibility, typicality, temporal, metadata-safety, and grounding criteria before they become audit-ready scenario packs.

Current strongest result

96 sequential hourly transitions.

Synset has validated acute-care trajectories across four days of hourly clinical evolution under trajectory-level certification criteria.

96

Sequential hourly transitions

Validated acute-care trajectory horizon.

97

State points

Initial anchor plus 96 transitions.

14,143

Accepted acute candidates

Accepted-only validator audit population.

0

Accepted audit violations

Observed among accepted acute candidates.

~0.021%

Upper bound

Conservative accepted-only audit estimate.

Caveat: Accepted-only validator audit, not held-out future patient accuracy, clinical truth, regulatory approval, or synthetic-control-arm validity.

Release certificates

Certified means passed release checks.

Certified means the artifact passed Synset's stated release checks. It does not mean the record is real, the future is guaranteed, or the output is a clinical recommendation.

Internal release record

Release Certificate

  • Clinical admissibilityPASS
  • Statistical typicalityPASS
  • Trajectory consistencyPASS
  • Metadata safetyPASS
  • Text groundingPASS

Clinical admissibility

Values and transitions satisfy stated clinical constraints.

Statistical typicality

Patterns remain consistent with empirical cohort-level structure.

Trajectory consistency

Patient state changes remain plausible across time.

Metadata safety

No source IDs, internal linkage, or forbidden identifiers.

Text grounding

Notes, dialogue, and claims-style artifacts stay grounded to structured evidence.

Evidence tiers

Generation, release, and future accuracy are separate claims.

Synset reports accepted candidates, release gates, withheld checks, self-consistency, and true held-out future validation separately instead of flattening them into one score.

Tier 1

Held-out future validation

Compares accepted synthetic continuations against real future observations that were hidden during generation.

Acute and chronic row-level future validation remain pending until sufficient observed future tables are available.

Tier 2

Independent / withheld validation

Tests accepted outputs against validators or empirical-prior constraints separate from held-out patient-future accuracy.

Acute: 0 / 14,143 accepted-only validator violations. Chronic: 0 / 3,953 final withheld-prior violations after release gating.

Tier 3

Release self-consistency

Checks whether regenerated continuations remain consistent with trajectory envelopes that passed release gates.

Measured internally for acute and chronic. This is a stability signal, not ground-truth accuracy.

Validation blind spots

Built to expose what retrospective validation can miss.

Synset is designed to produce evidence packs that can show when ordinary retrospective validation looks acceptable while scenario-controlled cohorts expose brittleness. Examples here are illustrative scenario templates unless backed by partner data.

  • missingness
  • delayed signal
  • care-pathway variation
  • subgroup presentation
  • documentation drift
  • longitudinal drift
  • counterfactual inconsistency
  • unsafe reassurance
  • failure to ask for missing context
  • contraindication misses
Benchmarking synthetic clinical data

Synset is both generator and evaluator.

Synset can benchmark whether a synthetic clinical dataset is clinically allowable, temporally plausible, text-grounded, and audit-ready.

  • admissibility scorecard
  • temporal consistency report
  • text-grounding report
  • subgroup/support report
  • release-certificate summary
  • downstream utility / robustness check where available
Research and roadmap

Roadmap directions stay labeled as roadmap.

These are roadmap or partner-driven directions, not current validated claims unless specifically marked as validated.

Roadmap

life-sciences and RWE exploration

Roadmap

trial feasibility exploration, not synthetic control arms as the current identity

Roadmap

imaging, genetic, and waveform sidecars

Roadmap

speech and dialogue simulation

Roadmap

enclave teacher / public student workflows

Request the evidence behind a scenario pack.

Synset can scope scenario packs, evidence reports, or model-under-test evaluations with explicit claim boundaries.

Request audit pilotView use cases