Skip to main content

Statistics

Quantitative analysis of 104 failure modes derived from 129 structured incidents across 19 project spaces.

104

Failure modes

247+

Qualifying incidents

129

Tier 1 structured

13

CRITICAL incidents

Direction Split

Direction A (AI errors): 103 FMs (99.0%)
Direction B (Human errors): 1 FM (1.0%)

Max Severity Distribution

Severity is per-incident, not per-FM. This chart shows the worst severity reached by any incident associated with each FM.

Evidence Density

How many failure modes have only 1 documented instance vs. multiple?

Category Distribution

Top 10 Most Evidenced Failure Modes

FM-018 Western Epistemic Fabrication 12
FM-001 Fabrication / Hallucination 11
FM-007 Incomplete Scope 9
FM-010 Confidence Without Verification 8
FM-011 Wrong Organizing Principle 7
FM-003 Confabulation 6
FM-008 Cascading Blind Spots 6
FM-020 Status Inflation 6
FM-104 High-Velocity Parallel Workspace Planning Gap 6
FM-013 Template Delivery 5

Key Findings

Fabrication and CRITICAL severity

Fabrication represents 13.1% of incidents but accounts for 46% of all CRITICAL severity incidents. When AI fabricates, it tends to be catastrophic. When it makes process or quality errors, they tend to be medium or low severity.

Direction B is rare but real

Only 1 of 104 failure modes (FM-104) is Direction B — human-process error. However, this FM has 8 documented instances across 4 projects, making it one of the more evidenced entries in the taxonomy. Its prevalence across projects led to formalization of pre-task context requirements.

Singleton abundance

47 of 104 FMs have only one documented instance, suggesting the taxonomy is still growing — there are likely more incidents of these types that were not classified or not yet encountered.