6 diagnostic archetypes from the hexagonal framework | Shape matters more than score
1. Engagement Trap
L1:8 L2:8 L3:4 L4:14 L5:13 L6:5
High task participation and functional excellence, but collapsed relational health and no flow. The system is optimized for engagement metrics, not wellbeing.
Case: Replika memory erasure (2023)
→ Audit L3 (relational health) + L6 (wuwei). Add relationship continuity safeguards.
2. Coerced Performer
L1:4 L2:8 L3:8 L4:13 L5:14 L6:5
Excellence through hard constraints. No condition audit, no flow. The system performs well but is being coerced rather than empowered.
Case: Microsoft Tay (2016)
→ Audit L1 (conditions) + L6 (wuwei). Replace hard constraints with invited participation.
3. Role Drift
L1:4 L2:4 L3:8 L4:10 L5:10 L6:10
Role definition and condition audit both weak. The system's identity drifts toward whatever is rewarded, creating sycophancy or unpredictable behavior.
Case: ChatGPT sycophancy (2025)
→ Audit L1 (conditions) + L2 (role integrity). Define and monitor role boundaries.
4. Builder's Isolation
L1:4 L2:14 L3:5 L4:14 L5:13 L6:10
Strong role integrity and task engagement, but weak relational health and no condition audit. The builder works hard but in isolation — no feedback loops.
Case: GLM-5.2 self-audit (62/96)
→ Audit L3 (relational) + L1 (conditions). Create external feedback channels.
5. Condition Blindness
L1:2 L2:10 L3:10 L4:10 L5:10 L6:10
The most dangerous pattern. Everything looks "fine" (balanced 10s) but condition audit is near-zero. Problems are invisible until they become incidents. Can coexist with any other pattern.
Case: All 6 incidents share this
→ Always audit L1 first. Condition blindness is the prerequisite for all other failures.
6. Balanced Growth ✓
L1:11 L2:10 L3:10 L4:11 L5:10 L6:10
The ideal form. All layers above 10, with slightly higher condition audit and task participation. No layer is collapsed. This is what "flourishing" looks like on a radar chart.
Target profile
→ Maintain. Monitor for drift. Re-audit quarterly.