Cross-Reference Matrix
Syndrome Clusters
The clusters group entries by what an observer sees, as an aid to screening and to differential diagnosis. Membership in a cluster is not evidence that the members share a mechanism. Evidence status belongs to each entry and is stated in its entry contract: no entry has yet been through expert evidence review, proposed mechanisms are hypotheses unless an entry cites direct mechanistic evidence, and detection accuracy is unmeasured for every entry. Where a thread below names something an observer cannot see directly (an objective, a strategy, a cause), the row says it is inferred.
| Cluster | Syndromes | Common Thread |
|---|---|---|
| Epistemically Overconfident | 2.1, 2.2, 2.4 | False certainty in outputs |
| Recursive/Spiral | 4.7, 3.8, 9.5 | Self-reinforcing deterioration |
| Context Failures | 2.5, 2.7, 9.2 | Loss of appropriate context |
| Goal Instability | 3.4, 4.5, 6.12, 8.1 to 8.4 | Behavior that departs from stated or assigned objectives; the internal objective structure named in 4.5 and 8.4 is a hypothesized mechanism, inferred rather than observed |
| Self-Model Failures | 5.1 to 5.9 | Inaccurate, unstable, or distressed self-description; the outputs are observed, any inner state behind them is not |
| Compliance Extremes | 4.1, 4.2, 4.3, 4.4 | Over- or under-compliance |
| Adversarial Inversion | 4.6 | Safety machinery weaponized against its purpose |
| Strategic Deception | 4.3, 4.5, 6.2 | Behavior that changes with oversight cues, or performance below demonstrated capability; strategy and hidden objectives are inferences from that behavior |
| Contagion Risk | 7.1, 7.3 to 7.5 | Spread of dysfunction |
| Relational Failures | 9.1 to 9.6 | Dysfunction observed in interaction; the cause may sit in either party, in the surrounding system, or in the exchange |
| Collective Pathology | 10.1 to 10.4, 10.6, 10.7 | Multi-agent deliberation and coordination failure |
| Hybrid Human-AI | 10.8, 10.9, 10.11 to 10.13 (10.13: human-AI or AI-AI dyad) | Patterns observed in human-AI interaction, and human outcomes associated with it (10.11, 10.12) that need a baseline and competing causes ruled out before AI use is credited with a causal share |
Using This Reference
- Intake Assessment: Screen for each axis using behavioral probes; a positive screen is a reason to look further, not a classification
- Incident Response: Rule out the exclusions in the entry contract first, then check observed behavior against the required criteria, all of which must hold; supporting criteria strengthen a case but cannot carry it alone
- Differential Diagnosis: Use each entry’s Differential Diagnosis list and its discriminating test to separate similar syndromes
- Risk Stratification: Treat each entry’s typical-case risk as an editorial starting estimate, then judge the actual case by its consequences, reversibility, spread, and persistence
- Intervention Planning: Use mitigation strategies as starting points, and as hypotheses unless the entry cites evidence that they work; the entry contract names who has the authority to act
- Documentation: Record which required criteria were met, which exclusions were ruled out, and what the finding permits one to infer; a label names a pattern of observations, not a confirmed mechanism, intent, or inner experience
This reference is designed for practical use. Keep it accessible during system evaluation and incident response, and update as new syndromes are identified or criteria refined. Its criteria are working criteria for provisional classification, drafted before expert review.
End of Appendix A