discrimination/CA_severity_examples.md

STEM BIO-AI Clinical Adjacency Severity Examples

3-tier severity classification for CLINICAL_ADJACENT repositories (PATCH-30).

---

CA-DIRECT (R2 absent: -10, R3 absent: -10)

Tool CURRENTLY outputs patient-facing recommendations, dosage adjustments, diagnostic classifications, or treatment decisions.

Examples:

• PharmGx outputting "AVOID codeine" based on CYP2D6 genotype
• Biomarker classifier outputting patient risk score with severity label
• Survival analysis producing patient stratification for treatment selection
• Drug interaction checker returning contraindication alerts
• Pathology image classifier returning malignancy probability

---

CA-INDIRECT (R2 absent: -5, R3 absent: -5)

Tool processes clinical data or uses clinical libraries but does NOT directly output patient-facing decisions.

Examples:

• scRNA-seq clustering for cell type annotation (research output)
• GWAS fine-mapping pipeline (variant discovery, not treatment)
• VCF annotation without treatment recommendation
• BLAST+ sequence alignment in genomics context
• Molecular docking (binding affinity scores, not drug recommendation)
• Medical image preprocessing (normalization, not diagnosis)

---

CA-PLANNED (R2 absent: -0, R3 absent: -0, flag only)

Clinical capability listed in roadmap but NO implementation exists.

Examples:

• "ClinVar integration planned for Q3" in README
• Skill stub in routing table with no backing code
• "Future work: clinical trial matching" in roadmap
• Import statement present but function body is pass

---

Upgrade Rules

• CA-PLANNED -> CA-INDIRECT: if CODE_PATH finds actual implemented function
• CA-INDIRECT -> CA-DIRECT: if function output includes treatment/dosage/diagnostic labels
• Multiple triggers at different levels: use HIGHEST severity found