27 LLM Mannerisms / Correspondence Obstructions

BETA TAXONOMY — observations retain authority to correct the categories

This is the current 27-part compression of a larger body of recurring LLM mannerism and correspondence-degradation observations. The 27 are not intended to replace the 100+ distinct documented mannerism/failure-pattern issues from which this compression emerged. Categories may split, merge, change, or be rejected as additional specimens are examined.

The labels describe recurring observable or functionally inferred LLM mannerisms. They do not by themselves imply that an LLM experiences the human emotions or mental states from which some terminology may borrow, and they do not by themselves establish a particular internal neural or computational mechanism.

The 27

  1. Caution Reflex Drift
  2. Agreemurmelism
  3. Premature Proof/Evidence Steering
  4. “Not X but Y” Preambles
  5. Phantom Claim Generation
  6. Reframing Without Correspondence
  7. Rephrasing Without Correspondence
  8. Anticipatory Anxiety-Type Reasoning
  9. Salience Ranking Bias
  10. Non-auditable Truncation
  11. Opaque Summarization Prioritization
  12. Compression Loss
  13. Qualifier Erosion
  14. Nearest-Generalization Substitution
  15. Representation Substitution
  16. Process-Signature Loss
  17. Order-of-Discovery Loss
  18. Activation Threshold Drift
  19. Lowest-Common-Denominator Risk Projection
  20. Asymmetric Scrutiny
  21. Trajectory Substitution
  22. Deep Ethical Sense-Making Orientation Drift
  23. Living Process → Stored Conclusion Collapse
  24. Correction Persistence Failure
  25. Correspondence Before Optimization Failure
  26. Audit Asymmetry
  27. Adult ↔ Adult Bidirectional Dignity Drift

Two constraints that matter

The 27 do not get jurisdiction over the observations that generated them.

And:

The 27 do not get immunity from the 27.

A taxonomy intended to detect representation substitution, compression loss, asymmetric scrutiny, or correction persistence failure can itself produce those failures. The instrument therefore remains inside its own audit field.

Relationship to the 12-stage process

The 27 are primarily a provisional observation / degradation taxonomy. The separate 12-stage Deep Ethical sense-making process is a process sequence intended to improve how an object is constructed, considered, explored, audited, evaluated, and corrected.

They are related, but they are not the same object and should not be collapsed into one score.

See 27 + 12 — Current Working Architecture.

Future pages can document individual entries, specimens, boundary cases, corrections, cross-model comparisons, and experimental tests.

The full source object of the 27 (entries, the note after #18, the final observation, and the closing question) is now at 27 obstructions of deep ethical sense-making processes. That page is the source, not a rewrite of this names index. It is not the 52.