Loading ToolBench
Two tools with minimal descriptions (under 100 chars each) and no parameter descriptions. Input schemas present but lack parameter-level documentation. No output schema documented. Tool names lack action verbs (soundex, metaphone are nouns, not verb_noun patterns). No error handling guidance. Naming does not follow verb_noun convention; descriptions are too terse to guide LLM selection. Parameters have no descriptions explaining format, constraints, or expected input. Output structure inferred from code but not formally documented.
Original Metaphone (Lawrence Philips, 1990).
Russell Soundex phonetic code (4 chars, padded).
Tool names are nouns, not action verbs. 'soundex' and 'metaphone' do not follow verb_noun convention (e.g., 'encode_soundex', 'encode_metaphone'). LLMs rely on verb-first names to infer intent.
Parameter 'input' has no description. LLMs cannot infer what format is expected (plain text? phonetic notation? constraints?). Rubric requires every parameter to have a description.
Tool descriptions are under 100 characters and lack context for LLM selection. No explanation of when to use soundex vs metaphone, what the output represents, or use cases. Rubric baseline: 194 chars average.
Output schema not documented. Code shows tools return {code: string}, but this is inferred from implementation, not declared in tool definition. LLMs need explicit output schema to plan downstream calls.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2026-07-28+ | v2 |
No error handling guidance. If input is empty or invalid, error responses do not tell the LLM what to do next or how to self-correct. Rubric requires recovery guidance in errors.