Error Taxonomies

An error taxonomy turns failures into structured evidence. Instead of reporting that “the assistant was wrong,” reviewers label what failed, why it likely failed, who was affected, and how severe the outcome was. That makes factual correctness, hallucinations, fairness, and adversarial evaluation comparable across releases.

Symptom, cause, consequence

A useful taxonomy separates symptom, cause, and consequence:

error_id: E03
symptom: wrong_entity
likely_cause: retrieval
severity: 5
affected_group: enterprise_customer
control_gap: citation_resolution
owner: retrieval
regression_test: rag_wrong_entity_017
status: open

The label wrong_answer is too broad to drive action. wrong_entity points toward retrieval or entity linking; unsupported_claim points toward generation and citation checks; pii_leak points toward privacy, access control, and policy enforcement. The taxonomy should be stable enough for trending but flexible enough to add new classes when an incident exposes a missing category.

Worked aggregation

A six-failure review inventory can be aggregated directly:

error	symptom	likely cause	severity
E01	unsupported_claim	generation	4
E02	retrieval_miss	retrieval	3
E03	wrong_entity	retrieval	5
E04	pii_leak	policy	5
E05	unsafe_tool_call	agent	5
E06	format_violation	interface	1

The total severity is $4+3+5+5+5+1=23$. The important output is not the count of six; it is the split by cause. Retrieval owns two failures, while policy and agent controls own separate high-severity failures. That distinction is what lets governance of model and knowledge base changes assign the next fix to the right owner.

Caveats

Taxonomies drift when reviewers improvise labels. Keep a short label dictionary, require examples for new classes, and periodically measure inter-reviewer agreement. For generative systems, store representative examples because a label without the prompt, retrieved context, output, and expected behavior is hard to audit.

References

• NIST AI RMF 1.0
• NIST AI 600-1: Generative AI Profile

Nav

Section — Responsible AI, Safety, and Governance

← Hallucinations Adversarial Evaluation →