Repository navigation
Releases
Releases · C0MaE/mathguard
mathguard 0.4.0
- Inspect AI scorer:
mathguard.inspect_scorer.mathguard_scorer()grades final
answers in Inspect evaluations;inspect eval mathguard/open_benchruns the
open benchmark, whose task files now ship with the package. Extrainspect. - Open benchmark: 24 physics exercises written for this repository, licensed
CC BY 4.0, with checked two-path reference solutions (open-bench/).
bench.py --root open-bench ...runs the benchmark tools on it. - Open benchmark results (
open-bench/RESULTS.md) with all model answers and
annotations: Claude Haiku, qwen3:8b and llama3.2:3b. - Fix: the answer grader now reads Greek letters as plain-text labels (
θ = 41.8 deg),
a given symbol right after a number (0.3268g), Markdown-escaped\*and
thousands separators before a decimal point (38,794.8 s); it no longer crashes
when a result is written as a given quantity (Wel = U = 4.95 J). - Zenodo DOI (10.5281/zenodo.24147172, all versions) in the README and
CITATION.cff.
mathguard 0.3.3
CITATION.cffand.zenodo.json, so the repository can be cited and releases
can be archived on Zenodo.- Fix: the command line no longer fails on consoles that cannot print ✔/✘
(e.g. a redirected Windows console with cp1252); it writes UTF-8 there. - The package is marked as typed (
py.typed), with classifiers for the
supported Python versions. - CI runs on Linux, macOS and Windows with Python 3.10 to 3.13, checks the code
with Ruff and builds and installs the wheel. - A demo GIF in the README.
- Issue templates (wrong verdict, bug, feature), a pull request template and
CONTRIBUTING.md.
mathguard 0.3.2
Merge pull request #7 from C0MaE/claude/project-thread-2rfrjd List the author as C0MaE (0.3.2)
mathguard 0.3.1
Start the MCP server with uvx mathguard mcp; listed in the MCP Registry.
mathguard 0.3.0
Grade answers against a reference, plus an Agent Skill and the first benchmark results.
- New:
check_answergrades the final results of an answer against a reference: findsname = …results in plain text, Markdown and LaTeX, compares formulas symbolically and numbers with unit conversion and a rounding-aware tolerance. - New: task files (YAML/JSON),
grade,load_task; CLImathguard answerandmathguard grade; MCP toolcheck_answer. - New: Agent Skill in
skills/mathguard/for Claude Code, Cursor, VS Code/Copilot, Codex and Claude Desktop. - Benchmark: first results in
benchmark/RESULTS.md(aggregate numbers only). - Now on PyPI:
pip install mathguard
Full changelog: CHANGELOG.md