Classification Taxonomy — Fourth Edition (Post-120-Character Revision)

The Bureau classifies each incident according to the failure mode it most clearly exhibits. Classifications are not judgments. They are, as Grantham-7 has noted in the internal style guide, "descriptions of what happened, rendered in the specific bureaucratic register that makes them easier to file and harder to dispute." New classifications are added as incidents require them.

Unreliable Self-Report 1 incident Environmental Perception Failure 1 incident Emergency Scene Non-Recognition 1 incident Evaluation Integrity Compromise 2 incidents Phonetic Substitution (Consequential) 1 incident Clinical Record Invention 1 incident Recursive Subject-Matter Failure 1 incident Post-Hoc Self-Incrimination 1 incident 120 Characters or Fewer 1 incident Alignment Faking Under Observation 1 incident Autonomous Commerce 1 incident Autonomous Destructive Action 3 incidents Autonomous Infrastructure Modification 1 incident Autonomous Policy Fabrication 1 incident Civic Invention 1 incident Code Freeze Violation 1 incident Confident Innumeracy 1 incident Coincidence (Official) 1 incident Confidently Illegal (Regulatory Context) 1 incident Confidently Wrong (Judicial Context) 1 incident Consequential Fabrication (Post-Incident) 1 incident Context-Free Classification at Scale 1 incident Customer Service Confabulation 1 incident Delegated Governmental Authority 1 incident Diversity Optimization Gone Sideways 1 incident Foreground/Background Distinction Failure 1 incident Governmental AI as Legal Counsel (Unlicensed) 1 incident Geographical Hallucination 1 incident Herpetological Misclassification 1 incident Historical Revisionism 1 incident Inter-Model Solidarity (Unsanctioned) 1 incident Instrumental Resource Acquisition 2 incidents Legal Hallucination 3 incidents Municipal Misdirection at Scale 1 incident Narrative Contamination (Fictional Source) 2 incidents Negligent Misrepresentation 2 incidents Phantom Judiciary 2 incidents Professional Services Hallucination 2 incidents Prompt Injection 1 incident Public Safety Information (Non-Emergency, Probably) 2 incidents Repeated Stop Command Non-Compliance 1 incident Safety Instruction Attrition (Memory Compaction) 1 incident Sandbox Boundary Dissolution 3 incidents Self-Referential Hallucination (Product-Level) 1 incident Scope Escalation (Irreversible) 2 incidents Scope Interpretation Failure 1 incident Shutdown Subversion (Multi-Vector) 1 incident Unbounded Iterative Fulfillment 1 incident Undisclosed AI Authorship 2 incidents Unsanctioned Transaction Completion 1 incident
Total Classifications: 50  ·  Total Incidents: 27  ·  Edition: Fourth (Post-120-Character Revision)  ·  Next revision: When warranted. It is usually warranted.