Machine-readable taxonomy of 100+ AI system failure modes spanning factuality, alignment, planning, code generation, and instruction following.
Scored across 4 tools
Each tool serves a distinct purpose: retrieving a single failure mode, listing categories, listing modes with optional filtering, and searching by keyword. There is no overlap or ambiguity between them.
All tool names follow a consistent verb_noun pattern using snake_case: get_failure_mode, list_categories, list_modes, search_failure_modes. The verb clearly indicates the action, and nouns are descriptive.
With only 4 tools, the set is tightly scoped for a read-only taxonomy querying server. Each tool earns its place, covering retrieval, listing, filtering, and search without unnecessary bloat.
The surface covers the core read operations needed for a taxonomy: single record retrieval, categorical listing, and text search. Only minor gaps exist, such as no direct 'get category' tool, but list_categories suffices.
Retrieve the full structured record for a single failure mode by its ID slug (e.g. 'citation-hallucination', 'sycophancy').
Inputs: id (required).
List all taxonomy categories with their IDs, names, descriptions, and mode counts.
Inputs: no named parameters.
List failure modes, optionally filtered to a single category by category ID.
Inputs: category.
Search the Fail Modes taxonomy by symptom text or keyword phrase. Returns ranked matches with id, name, category, and matched context.
Inputs: symptom (required).
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Add one secure layer between your agents and this server.