We evaluated a two-component clinical AI system across 125 rheumatology scenarios in seven protocols. The first component generates a candidate clinical response and subjects it to a four-dimension verification loop (clinical accuracy 0.30, safety 0.30, therapeutic management 0.20, resource stewardship 0.20). Responses below threshold receive specific corrective feedback and are regenerated (maximum three cycles). The second component applies PCA to 81,502 rheumatology article embeddings, assigning 6-bit precision to dimensions 1-128 (68% variance, encoding diseases and treatments), 4-bit to dimensions 129-512 (25%, comorbidity patterns), and 2-bit to the remainder (7%). This compresses the index from 335 MB to 39 MB while preserving 95% recall at 10 passages. Generic random rotation achieves only 87% because it destroys the anisotropic variance structure of specialist medical language. Results: combined system scored 8.90 vs 8.18 unaugmented GPT-4o, reduced hallucination from 12-15% to under 2%, compressed inter-scenario variance by 89%, improved safety by 7.3 points and escalation appropriateness by 10.0 points. Critical finding: naive retrieval degraded performance below baseline (7.92 vs 8.38 in Protocol B). The same architecture with corpus-curated compression resolved the paradox (Protocol G). Verification and retrieval are jointly necessary and individually insufficient.