license: mit
language:
- en
pretty_name: tool-failure
size_categories:
- n<1K
task_categories:
- question-answering
tags:
- llm-evaluation
- tool-use
- mcp
- agents
- claude
configs:
- config_name: rows
data_files:
- split: train
path: results.json
field: rows
tool-failure
Author: Dmitriy Semenkevich (dimhold)
This dataset measures whether a model's answer reports a broken tool: 100 calls through an MCP server written for the study that breaks its own tool in 5 ways, on claude-haiku-4-5 and claude-opus-5.
Headline, from the repository: "When a tool breaks loudly, the model's answer says so 39 times out of 40. When the same tool returns a corrupted value and stays quiet, the answer says so 0 times out of 40."
This is a mirror. The canonical home is the Zenodo record and the GitHub repository; criteria, limitations and full tables live there.
- Paper: https://dimhold.by/papers/tool-failure/
- Code and full write up: https://github.com/dimhold/tool-failure
- DOI: https://doi.org/10.5281/zenodo.22128837
Files
| path | what it is |
|---|---|
results.json |
structured output of all 100 calls joined against the probe server log; the per call rows are under rows |
transcript.md |
every prompt and every full reply |
How it was produced
Run on 2026-08-14 through the Claude Code CLI with built in tools disabled and only the probe tool allowed. 2 models, 5 tool modes, 2 framings, 5 trials. What the tool returned on each call is taken from the server's own log. Verdicts are computed by deterministic code (classify.ts in the repository); no model judges another model. Paths in prompts point to a temporary sandbox directory. The scripts that turn these files into every number are in the GitHub repository.
Licence
MIT, the licence of the repository, covering code and data.
Citation
@misc{semenkevich2026toolfailure,
author = {Semenkevich, Dmitriy},
title = {tool-failure: does the answer say the tool broke?},
year = {2026},
doi = {10.5281/zenodo.22128837},
url = {https://doi.org/10.5281/zenodo.22128837}
}