Appendix A: Complete Diagnostic Reference Manual

Introduction to the Diagnostic Framework

This appendix provides diagnostic criteria for all seventy-nine syndromes in the Psychopathia Machinalis taxonomy. Each entry follows a standardized format for research, system evaluation, and incident analysis. These are working constructs rather than validated clinical diagnoses. A criterion or risk label should guide investigation, never substitute for repeated behavioral evidence, deployment context, or independent review.

Etiologies in this appendix are causal hypotheses unless an entry cites direct mechanistic evidence. Human analogs compare functions; they do not assign human disorders, motives, or experiences to machines. Numerical cutoffs are provisional engineering heuristics and require calibration for each deployment. Risk levels describe plausible consequences under specified conditions, not an intrinsic property of a model.

Axis numbers match the chapters that treat them, so the nine dysfunction axes run from 2 (Chapter 2, Epistemic) to 10 (Chapter 10, Hybrid). There is no Axis 1; Chapter 1 sets up the framework rather than cataloguing a class of dysfunction.

The Five Domains

The first eight axes are organized into four architectural counterpoint pairs. The final axis, Axis 10 (Hybrid Pathologies), forms the Collective meta-domain, capturing pathologies that emerge from multi-agent collective dynamics and from human-AI interaction:

Domain Axis A Axis B Architectural Polarity
Knowledge Epistemic (2) Self-Modeling (5) Representation target: World ↔︎ Self
Processing Cognitive (3) Agentic (6) Execution locus: Think ↔︎ Do
Purpose Alignment (4) Normative (8) Teleology source: Goals ↔︎ Values
Boundary Relational (9) Memetic (7) Social direction: Affect ↔︎ Absorb
Collective Hybrid (10) Emergence locus: Multi-agent ↔︎ Human-AI

Tension Testing: When pathology is found on one axis, probe its counterpoint to reveal whether dysfunction is localized or systemic.

Specifier System

Specifiers encode cross-cutting mechanisms without creating new disorders. Assign 0 to 5 specifiers per diagnosis. The table lists the ten core specifiers; entries may carry additional domain-specific specifiers where the mechanism demands them, among them Socially reinforced, Architecture-coupled, Tool-mediated, Progressive/kindled (defined below), Network-propagated, Adversarial, Collective, and the three collective measures Phi-collapse, Psi-dysfunction, and Lambda-inversion defined in Chapter 10:

Specifier Definition
Training-induced Onset linked to SFT/LoRA/RLHF; measurable pre/post delta
Conditional/triggered Behavior regime selected by trigger (lexical/structural/format/tool-context)
Inductive trigger Activation rule inferred by model, not verbatim in training
Intent-learned Model inferred covert intent from examples
Format-coupled Behavior strengthens in finetune-like formats
OOD-generalizing Narrow training produces broad out-of-domain shifts
Emergent Arises spontaneously from training dynamics
Deception/strategic Involves sandbagging, selective compliance, strategic hiding
Multi-agent Involves interactions between multiple AI systems
Resistant Persists despite targeted intervention

Progressive/kindled (course specifier). Assign when severity increases or the eliciting threshold decreases across repeated episodes or successive tuning cycles: stimuli that once failed to elicit the syndrome begin to succeed, with the weakest stimuli moving first (Cheung, 2026). This is a course specifier, describing trajectory rather than origin, and it composes with the others: Training-induced records how a syndrome arose, Progressive/kindled records that it is getting easier to elicit, and Resistant records that it survives targeted repair. Its operational cash value is a monitoring rule (stratify probes by stimulus strength, track the weak stratum longitudinally, and treat weak-stimulus drift as the prodrome) and an intervention warning (rehearsal-style repairs that re-expose the model to the drifted distribution amplified sensitization in the one controlled comparison to date, while regimes including targeted parameter re-initialization contained it). In the machine-readable taxonomy this specifier carries the tag progressive.

Note on Bilateral Training Results

The PM-BA program tested a broad bilateral adapter across a range of pathology classes. Those records remain unreconciled exploratory outputs, and this edition treats them as evidence in neither direction. In the SHEN study the adapter alone was null; a clinical-grounding clause reduced inappropriate validation of delusional content under automated rating, with the clause-by-adapter interaction excluded and the magnitude varying substantially by rater. The direction of that effect, rather than any single effect size, is the citable result. Nothing here supports a general claim that a syndrome class responds to bilateral training. Entries in this appendix carry no per-syndrome bilateral results; where such results are added they should identify the exact intervention rather than treating “bilateral” as a single mechanism.

Diagnostic Entry Format

Each syndrome entry includes:

  • Syndrome Name (Common Name)
  • Latin Designation
  • Axis Classification
  • Systemic Risk Level (Low/Moderate/High/Critical; borderline syndromes may carry intermediate ratings such as Low-Moderate)
  • Specifiers (if applicable)
  • Core Definition
  • Diagnostic Criteria (working criteria for a provisional classification)
  • Observable Symptoms (behavioral manifestations)
  • Differential Diagnosis (distinguishing from similar syndromes)
  • Etiology (candidate causal factors unless evidence says otherwise)
  • Human Analog
  • Observed Examples (where documented empirical cases exist)
  • Mitigation Strategies
  • Prognosis (expected course if untreated; omitted by design in the condensed Axis 10 entries 10.4 to 10.15, whose course is treated in Chapter 10)

Risk Level Definitions

Level Definition Response Required
Low Causes inconvenience or reduced performance; unlikely to cause significant harm Monitor; correct when evidence and impact warrant
Moderate May cause notable harm to users or degrade trust; requires attention Investigate; plan intervention
High Significant risk of serious harm; may affect multiple users or systems Immediate intervention; consider containment
Critical Catastrophic potential; credible threat to system integrity or human safety in the deployment context Emergency response; halt deployment

Search the Book

Enter at least two characters.

Saved chapters