v1.1.0

Added

• arquimedes fetch downloads the corpus from Hugging Face into the user data directory and builds a lean SQLite index. uvx arquimedes-mcp fetch is the same download, and uvx arquimedes-mcp serves that copy. The PyPI package is arquimedes-mcp and does not contain the problems.
• server.json is the official MCP Registry manifest for io.github.richer-richard/arquimedes.

License

• Problem statements, solutions, answers, and diagrams are no longer described as CC BY 4.0. Upstream terms are documented per dataset in docs/datasets.md. CC BY 4.0 remains only for the Mohs calibration, the category labels assigned by this project, and data/index.json.
• Each Hugging Face record carries its holder's notice. distribution is null where no open license was found. The copy is richard01172000/arquimedes.

v1.0.0

First stable release. From 1.0.0 on, the record schema, MCP tools, CLI, SQLite tables, and ArquimedesEngine methods only change in backward-compatible ways until 2.0 (see "Stability (1.x)" in the README). Full details are in CHANGELOG.md.

Breaking changes from 0.3.x

• MCP tools renamed: arquimedes_get_stats, arquimedes_search_problems, arquimedes_sample_problems, arquimedes_get_problem. The old names still work in tools/call but are not listed.
• answer.letter only holds the choice letter of a multiple-choice problem. 1,830 core free-response records that copied their answer there now have a null letter; answer.value is unchanged.
• source.number is an int when purely numeric (202 aops records), and get_problem returns the same record shape with a full, lean, or missing database.

Fixed

• The MCP server no longer exits on bad input; errors come back as tool results or JSON-RPC errors.
• Sampling is uniform over all matching problems instead of the first few hundred matches.
• Data fixes: 2 placeholder solutions removed, 25 HMMT titles repaired, 6 imputed difficulty notes nulled, one HMMT answer recovered from its solution sketch. Written solutions: 797,671.

Added

• Tests for the engine (all three database modes), MCP server, and CLI, run in CI with Python 3.10 and 3.14.
• A (dataset, category) index and ANALYZE at build time; filtered sampling takes under 200 ms.

Problem data is licensed CC BY 4.0. Code in arquimedes/ and tools/ is MIT. difficulty.mohs is a calibration across datasets, not a measured difficulty.

v0.3.2

Patch release for MCP clients.

• arquimedes-mcp now answers resources/templates/list with an empty list instead of a -32601 "Method not found" error. Clients that probe for resource templates, such as the MCP Inspector and Glama's inspector, no longer report an error on connect.

No data changes. Version 0.3.1 was skipped because Glama already used that number for its auto-release of v0.3.0.

v0.3.0

First tagged release of Arquimedes.

Data

• 1,038,963 problems across 9 datasets (numina, deepmath, bigmath, math, core, omnimath, olympiadbench, cemc, aops), with 797,672 written solutions, 277,002 model-reasoning traces, and 1,219 diagrams.
• Every record has difficulty.mohs on a 1.0–10.0 scale. It is a calibration across datasets, not ground truth.
• Rule-based classification brought uncategorized down from 376,384 records to 159. A category means a rule matched, not that a person checked it.

Tools

• arquimedes CLI: stats, search, sample, show, build, splits.
• arquimedes-mcp: a stdio MCP server with four read-only tools, also listed on Glama.
• arquimedes build --lean builds a 0.8 GB search index that reads solutions from the JSONL shards; the full build is about 1.7 GB.
• arquimedes splits exports a decontaminated train split (972,458 problems) and eval split (10,533).
• Web explorer: https://richer-richard.github.io/arquimedes/

Problem data is licensed CC BY 4.0. Code in arquimedes/ and tools/ is MIT.