Public note
acute-trajectory-methodology-v1.0
Public technical evidence note
Acute trajectory methodology.
This note documents the retained evidence behind Synset's 96-hour acute trajectory claim. It separates candidate acceptance, a post-certification certificate recheck, and the held-out evidence that has not yet been produced.
Evidence snapshot
Artifact snapshot
June 16, 2026
Maximum checked horizon
96 sequential hourly transitions
Summary schema
privion.final_longitudinal_post_certification_fidelity_metrics_summary.v1
Summary SHA-256
128ca95db9dbe036160b5d5e97738f7b6e85794b9a4f560b0754676a740b333d
The retained artifacts classify the configuration as internal and non-production. The 96-hour result is a support-checking result, not a forecast of a patient's future.
Candidate funnel
The retained summary combines four experiment arms at 64 and 96 hours. It records 36,917 generated candidates, 14,143 accepted candidates, and 22,774 rejected candidates. The accepted count is therefore not a count of 96-hour candidates alone.
| Horizon | Experiment arm | Generated | Accepted | Rejected |
|---|---|---|---|---|
| 64 hours | Re-anchor every 8 steps | 1,612 | 1,536 | 76 |
| 64 hours | Strict stop | 415 | 320 | 95 |
| 96 hours | Direct rollout | 24,000 | 1,535 | 22,465 |
| 96 hours | Re-anchor every 4 steps | 10,890 | 10,752 | 138 |
| Total | 36,917 | 14,143 | 22,774 | |
Acceptance record
The summary preserves the accepted flag from the source certificate table. Recorded rejection categories include categorical-count coherence, support-prior violations, and decoded hard or fatal rules. Categories can overlap, so this note does not present them as a mutually exclusive rejection breakdown.
The exact acceptance predicate and its validator and policy versions are not embedded in the retained public summary.
Accepted-only recheck
A separate post-processing step rechecked the 14,143 accepted certificate rows for certificate completeness, support coverage, decoded hard or fatal violations, quantile and absolute-delta support violations, typicality, and time monotonicity. No accepted row failed those checks in this snapshot.
This is not independent of acceptance in a claims-grade sense. It uses the accepted certificate rows and overlapping certificate fields, not an independent dataset or held-out patient futures.
Statistical boundary
What the zero-event recheck means.
For zero observed recheck events among 14,143 accepted rows, the retained calculator reports an exact one-sided 95% Clopper-Pearson upper bound of 0.0002118, or about 0.0212%.
That bound applies only to the accepted-only certificate recheck in this snapshot. It is not a bound on future clinical error, downstream model failure, patient outcomes, or held-out trajectory accuracy.
Provenance fields not retained publicly
- Not recorded in the current public evidence packet: an embedded run timestamp; the date above is the retained artifact snapshot timestamp.
- Not recorded in the current public evidence packet: a public generator or model version.
- Not recorded in the current public evidence packet: a public validator version.
- Not recorded in the current public evidence packet: a public constraint or acceptance-policy version.
- Not recorded in the current public evidence packet: a public source-data or source-prior manifest.
Reproducibility
The public note exposes the retained counts, schema version, statistical method, and summary hash so the published snapshot can be checked for change. The underlying certificate rows, source data, and full regeneration inputs are not included here, so this page alone does not provide full reproduction of the run.
Claim boundary
What this evidence does not establish.
future patient accuracy
clinical truth or clinical utility
model safety
clinical or regulatory readiness
synthetic control-arm validity
customer-model hardening
generalizability outside the tested configuration