Loading ToolBench
Measurement datasets and reproducible methods on browser tools, OCR pipelines and AI-assisted development. Every dataset has a documented method and a control run.
Four tools with complete input schemas and clear descriptions. Naming follows verb_noun convention (list_, get_, extract_). Descriptions are well-written (100-200 chars, within baseline 194 avg). However, output schemas are entirely undocumented, LLMs cannot predict response structure. No parameter descriptions for optional/conditional params (e.g., 'name' in get_method). No error handling guidance. No pagination support despite tools returning lists. Missing tool annotations (readOnlyHint present in metadata but not in schema). Tool composition is sound but output documentation is a critical gap.
Read the citation a web page declares about itself — authors, title, journal, year, DOI, licence — and return it as a structured record with a ready-to-import RIS entry and BibTeX. Where the page is a paywall, an error page or a bot check rather than a work, it says so and refuses instead of inventing a reference: check the warning before using the fields.
Fetch one measurement dataset as JSON. Pass the dataset URL or filename returned by list_measurements. Contains the measured values, the control run and the conditions under which they were obtained.
Fetch a reproducible method published here as a skill: how to read a browser extension's permissions, how to measure OCR recall with a control run, or how to choose between print-to-PDF and screen capture. Call without arguments to list the available methods.
List the measurements published on this site, each with its page URL, the raw dataset URL where one exists, and what was measured. Start here.
Output schemas completely undocumented. LLMs cannot predict response structure (fields, types, nesting). All four tools return text content but downstream field names are opaque.
Parameter 'name' in get_method lacks description. Is it case-sensitive? What happens if omitted? Current description only lists enum values, not the parameter's role.
No error handling guidance. Tools can fail (network, 404, invalid JSON) but responses provide no recovery hints. E.g., 'No method named X' is good, but what if the endpoint is down?
list_measurements returns unbounded list of measurements. No pagination (limit, offset, next_cursor) or result count. Large datasets will blow context window.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Tool annotations (readOnlyHint) declared in metadata but not in inputSchema. Schema should include 'readOnlyHint: true' for all four tools to signal safety to LLMs.