Model console

Versions, drift, and how often the model is told to keep quiet.

v2.4.1In production
4.8Mean absolute error, mmHg
31.4%Insights gated out by design
8,412Nightly refits
Registry
VersionApproachDeployedMembers MAECoverageDriftStatus
Cohort-level feature importance SHAP, challenger model

Population-level attribution. It informs content priorities — it is never shown to a member as if it were their own result.

Why insights get withheld

Roughly a third of candidate insights never reach a member. That number going down without a model change would be the thing to worry about.

Human review queue 3 waiting
Unusual effect size — BL-4402
Alcohol coefficient 8.9 mmHg/unit, 4.1σ above cohort. Possible data-entry error.
New recommendation phrasing — R-028
Awaiting clinician sign-off before it can be published.
Drift warning — sodium proxy
Distribution shifted after questionnaire v4.2. Refit recommended.