Model Validation & Governance

Validation Baseline

Current release uses stratified simulation validation across workflow-specific cohorts and operating points.

• Calibrated probability outputs with confidence signals in every score response.
• Deterministic model versioning with per-request model_version visibility.
• Regime-aware denial routing with explicit distribution_profile controls and response routing metadata.
• Workflow-scoped details under details.denial, details.prior_auth, and details.reimbursement.
• Runtime lock checks for model artifacts, calibration snapshots, and threshold snapshots before API startup.

Governance Controls

• Drift monitoring and downgrade alerting with structured scoring telemetry.
• Support for explicit model_version pinning and fallback controls.
• Readiness endpoint /readyz for dependency and startup health checks.
• Prometheus metrics are available only on the private production network; detailed diagnostics require operator authorization.
• Incident-first posture: if material degradation is detected, deployments roll back to known-good locked artifacts.

Current Validation Evidence (2026-09-02 UTC)

Validation basis: Healthcare model quality is evaluated on stratified simulated claim cohorts; this is not retrospective real-world claims validation.

Validated release/runtime: version 1.0.765 on IONOS production using the bounded production single-seed profile (seed 42).

Current model versions: denial 0.3.3; prior authorization 0.2.3-prior-auth-on-file; reimbursement 0.2.1; appeals generator 0.1.2.

Current production functional validation

Current production bounded single-seed run — MISS: predictive scoring completed 0/140 requests; 140 returned HTTP errors.

• Denial requests completed: 0/40.
• Prior-authorization requests completed: 0/60.
• Reimbursement requests completed: 0/40.
• Appeals functional criteria: 8/8 in this single-seed run.

Predictive scoring returned HTTP 503 while recording feedback request IDs. Model-quality metrics from this runtime run are therefore not assessed.

Latency gate NOT APPLICABLE: observed overall p95 327.77 ms; published gate < 500 ms. The p95 includes failed predictive scoring responses, so it is not eligible for a latency PASS.

Locked model-artifact evaluation metrics

These statistical metrics come from the locked current model artifacts, not from the incomplete production functional run above.

• Denial: ROC AUC 0.7937; average precision 0.4570; Brier score 0.0847; selected calibration ECE 0.0032.
• Prior authorization: ROC AUC 0.9667; average precision 0.9443; Brier score 0.0069; ECE Not currently produced.
• Reimbursement: R2 0.8270; MAPE 0.1236.

Limitations: Simulation evidence does not establish performance on retrospective real-world payer claims.

Security Overview

• API auth is enabled by default with bearer token enforcement.
• API key lifecycle is managed through authenticated customer workspace flows.
• TLS is enforced at the public HTTPS ingress; backend services communicate on the private production network.
• Public health and readiness probes return minimal availability state; Prometheus metrics remain private and detailed diagnostics require operator authorization.

Reference

For API integration details and endpoint behavior: /portal/docs-home

For machine-readable schema: /openapi.json