v1.0.1

The parser reads the whole row, and the measurement is corrected.

This release ships the strict whole-row parser, the corrected coverage measurement, and the mobile inspector.

Fixed

• A row ends at its real punctuation, so a trailing comma or period no longer hides the rest of the line.
• A row with several top-level clauses is split into claims, and a relation inside a group becomes a named raw node.
• Accented lowercase calls such as \hat{l}(f, y), powered calls such as d^2(f, g), and conditions inside expectations are read as written.
• A trailing \text{...} is an annotation, not a parse failure.
• samosa --version and samosa --help work before a command, and the report version is 1.0.1.

Measured

On 36 recent arXiv papers:

• 1416 display math rows, of which 1219 make a claim.
• 686 claim rows parse end to end: 56.3%.
• 304 claim rows have no named unknown construct: 24.9%.
• That is 21.5% of all display rows.

The gate G3 asks for 70% parse coverage. It is not met yet. The method, the failure buckets and the correction are in docs/COVERAGE.md.

App

• On a phone, the inspector is a bottom sheet, and the header actions live in the overflow menu.

Install

npx samosa-math check paper.tex
npm i -g samosa-math

The MCP server is npx -y -p samosa-math samosa-mcp. The GitHub Action is Hussain004/SAMOSA@v1.

Full changelog: CHANGELOG.md

v1.0.0

This release is kept for history. See v1.0.1 for the corrected parser and the corrected coverage measurement.

The first release.

The engine checks every step of a derivation, exactly, over a finite field. No AI is in the verdict path, and nothing leaves the browser.

The engine. A tokenizer and a Pratt parser for the LaTeX people actually write. Shape inference with a union find over the dimension variables. Exact mode over GF(2^31-1) and GF(2147483629), with the bound computed per size and shown on the card. Gradients and Hessians by dual numbers. Exact expectations by rational quadrature. Implications by constraint completion. Counterexamples shrunk until they are small enough to read. Repairs by mutation search, residual fingerprint, missing assumption and shape.

The surfaces. The web app, the CLI (pretty, JSON, SARIF, GitHub), the MCP server, the GitHub Action, the Agent Skill, and the npm package.

npx samosa-math check paper.tex

Measured. The measurement first published here was lenient and is superseded. The current measurement of the latest source, on 36 recent arXiv papers, is 56.3% of claim rows parsed end to end and 24.9% read with no named unknown construct, 21.5% of all display rows. The method and the correction are in docs/COVERAGE.md. The parser and the corrected number ship in v1.0.1.