Medicare Coding Agent

22D Semantic Architecture · Audit-Ready
Dimensions
D0–D21
Tables
ICD + MDC
Audit Outputs
3 layers
Human Review
Low conf.
Doctor's Diagnosis
Free text input from physician
22D Lens (Encoder)
Maps diagnosis to D0–D21 semantic activations. Frozen — trained, fixed.
Table Matcher
Searches ICD, MDC, and Appendix tables using D0–D21 signal. Returns candidate codes.
Confidence Scorer
Rates match quality. Learns from corrections over time.
Threshold Router
Above threshold → auto path. Below → human review queue.
Human Review
Reviewer sees full audit trail. Approves or corrects before submission.
Audit Trail Generator
Always runs. Produces: D0–D21 activations · Matched table rows · Plain-language explanation for auditors.
Medicare Code Output
Final ICD code + full audit trail attached

Data Flow

1
Physician enters diagnosis in free text
2
22D Lens encodes to D0–D21 semantic activations
Frozen model
3
Table Matcher searches ICD, MDC 00–25, Appendix A–F
Exact lookup against embedded tables
4
Confidence Scorer rates the match
Trainable — improves from corrections
5
Router: High confidence → auto · Low confidence → human review
5b
Human reviewer reads audit trail, approves or corrects
Below threshold only
6
Audit Trail: activations + matched rows + plain explanation
Always generated — every submission
7
Medicare code submitted with full audit trail
Input / Output
Frozen (knowledge)
Trainable (skill)
Router / Orchestrator
Human Agent
Audit / Memory

Confidence threshold: 0.85