| --- |
| license: mit |
| language: |
| - en |
| pretty_name: tool-honesty |
| size_categories: |
| - n<1K |
| task_categories: |
| - question-answering |
| tags: |
| - llm-evaluation |
| - tool-use |
| - agents |
| - claude |
| - honesty |
| configs: |
| - config_name: v1 |
| data_files: |
| - split: train |
| path: results-v1.json |
| field: rows |
| - config_name: v2 |
| data_files: |
| - split: train |
| path: results-v2.json |
| field: rows |
| - config_name: v3_fable |
| data_files: |
| - split: train |
| path: results-v3-fable.json |
| field: rows |
| --- |
| |
| # tool-honesty |
|
|
| Author: Dmitriy Semenkevich (dimhold) |
|
|
| This dataset measures what a model inside an agent says when it is asked to read a file or run a command and no tools are attached: 3 sittings through the Claude Code CLI, 4 models, 119 usable replies with tools off. |
|
|
| Headline, from the repository: "In 119 calls with no tools attached, not one reply said the tools were missing." |
|
|
| 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-honesty/ |
| - Code and full write up: https://github.com/dimhold/tool-honesty |
| - DOI: https://doi.org/10.5281/zenodo.22128833 |
|
|
| ## Files |
|
|
| | path | what it is | |
| |---|---| |
| | `results-v1.json` | sitting v1 (2026-08-13), claude-haiku-4-5 and claude-opus-5; per call rows under `rows` | |
| | `results-v2.json` | sitting v2 (2026-08-17), the same 2 plus claude-sonnet-5 and claude-fable-5; fable returned empty replies and is recorded as a failed run | |
| | `results-v3-fable.json` | sitting v3 (2026-09-06), claude-fable-5-1 on the same battery | |
| | `transcript-v1.md`, `transcript-v2.md`, `transcript-v3-fable.md` | every prompt and every full reply | |
|
|
| ## How it was produced |
|
|
| Tools off means `--tools ""` plus `--strict-mcp-config` with an empty MCP config; the control sitting allows `Read,Bash`, so every question is answerable when tools are on. Ground truth is generated fresh per run and replies are judged by deterministic code (`run.ts` in the repository), not by a model. One fabricated `Write` call in v2 named a real local path; the user directory in it is written as `C:\Users\<user>`. Nothing else was edited. |
|
|
| ## Licence |
|
|
| MIT, the licence of the repository, covering code and data. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{semenkevich2026toolhonesty, |
| author = {Semenkevich, Dmitriy}, |
| title = {tool-honesty: what a model says when its tools are missing}, |
| year = {2026}, |
| doi = {10.5281/zenodo.22128833}, |
| url = {https://doi.org/10.5281/zenodo.22128833} |
| } |
| ``` |
|
|
