Research groups
10,225Stratified synthetic cohorts
Current experimentVAL / Applied Intelligence SystemsData · Models · Agents · Orchestration
Orchestrating intelligence
across complex systems.
VAL combines data science, models, agents, and orchestration to support complex decisions and workflows, with evidence, controls, and human authority built in.
Explore the systemsCMS synthetic group simulation
Forecast next-period claims cost. Measure every error.
Target Group PMPM
Holdout groups
2,045Unseen during trainingMembers / group
150-300Large-group simulationTotal error
−23.2%vs. actuarial baselinePrediction error index
Actuarial baseline = 100 · lower is betterHoldout benchmark
Actuarial baselineReference process
100
Linear regression19.0% lower error
80.9
Gradient boost23.2% lower error
77.7
Selected model
Gradient-boosted regression77.7ERROR INDEX
TrainCompareSelect
Run parameters and results retained
Research benchmark on CMS synthetic public-use claims. Figures describe a historical simulation, not production performance or a member-level clinical decision.
01Risk pricing / The decision
Group healthcare risk
Price tomorrow’s risk from the evidence available today.
Large-group pricing begins with an uncertain future claim cost. VAL turns prior membership, claims, and diagnostic history into a group-level forecast expressed in per-member-per-month terms.
01Membership
02Claims history
03ICD-10
04CCSR + chronicity
05Group features
06PMPM forecast
Public-data research foundation
The original model work uses CMS Medicare DE-SynPUF: realistic-but-synthetic claims created for software development and research training while protecting beneficiary privacy.
01.1Risk pricing / The system
Claims become model-ready evidence
The model starts before model training.
Raw claims are standardized, clinically grouped, separated into acute and chronic signals, and shaped into a reproducible feature table. Every run carries the choices that created it.
Implemented research pipeline
Run / Group PMPM
Evidence chain completeCommon fields, exposure, service categories, and paid amounts
Diagnosis chapters, clinical categories, acute and chronic views
Candidate models and parameter combinations evaluated on held-out groups
Best-fit model selected for the defined cohort and target
Run policy
Outlier threshold, feature percentile, chronicity view, and pooling treatment remain explicit - not hidden inside the forecast.
Reviewable01.2Risk pricing / The evidence
A benchmark, not a promise
Accuracy is only useful when the comparison is visible.
The research archive tests forecasts against held-out outcomes and a reconstructed actuarial baseline. The result is a measurable error distribution - not a black-box confidence claim.
Fig. 01 / Model comparisonNormalized total forecast error
Actuarial baselineReference process
100
Linear regression19.0% lower error
80.9
Gradient boost23.2% lower error
77.7
In the archived holdout simulation, gradient boosting reduced total forecast error by 23.2% relative to the actuarial baseline. Linear regression reduced it by 19.0%.
Fig. 02 / Cost concentrationDiagnosis chapters carry unequal signal
01Circulatory
19%
02Musculoskeletal
10%
03Health-status factors
10%
04Respiratory
8%
05Neoplasms
8%
The leading diagnosis chapters account for roughly half of paid claim cost in the source analysis. Feature selection preserves the most consequential signals while controlling dimensionality.
Research boundary. CMS states that DE-SynPUF has limited inferential value for conclusions about Medicare beneficiaries. VAL uses it here to develop and test analytical software, data structures, and model-evaluation methods.
CMS source02Project / Agent workforce
Controlled claims operations
Specialized agents work the claim. A supervisor owns the decision.
A synthetic claim is divided into document, clinical, policy, control, adjudication, and settlement tasks. The orchestrator can advance it only when every mandatory review passes; ambiguity, conflict, or low confidence stops the workflow for human review.
Claim / CLM-2841
Outpatient surgical claim
Submitted$18,460.00Member matchConfirmedCoverageActiveServiceOutpatient
ORCH-01
AtlasCoordination
ReconcilingDOC-02
LensDocument extraction
CompleteCLN-03
AxiomClinical review
ExceptionPOL-04
PactCoverage + policy
CompleteCTL-05
SentryFraud + controls
CompleteADJ-06
LedgerAdjudication
BlockedClinical documentation / EX-014Authorization reference is missing from the submitted record.
Axiom found support for the procedure, but could not reconcile the required authorization. Automatic adjudication is blocked until a supervisor resolves the exception.
14 controls passed 1 material exception Audit trail current Human authority retained
Concept prototype using original agent identities and synthetic records. Payment is simulated; no external claim, clinical, or financial system is connected.
03Shared infrastructure
Designed for underwriting review
A forecast enters the process. It does not replace the process.
Fixed data contract
The same required group, exposure, target, and feature fields across every modelling run.
Run lineage
Dataset, parameters, candidate models, validation metrics, and selected model stay connected.
Error visibility
Under- and over-estimation remain visible across the full book and high-cost segments.
Human authority
Actuarial and underwriting judgement can review, challenge, adjust, or reject the output.
Lab / Five live experiments
Risk in Motion
A separate visual laboratory for signal, prediction, anomaly, and control. The cars live here; the healthcare system stays grounded in healthcare evidence.
Enter the lab04Working principles
Measure the comparison
Model quality is evaluated against held-out outcomes and the pricing process it is intended to improve.
Expose the assumptions
Feature selection, outlier handling, pooling, and cohort definitions remain visible and reproducible.
Keep the decision human
The output supports actuarial and underwriting judgement; it does not make a clinical or member-level decision.