Extract formal math specifications from documents

DOC2MATH extracts a formal Mathematical Problem Specification (MPS) from a document, research excerpt, or word problem, under a strict "Zero-Inference Protocol": nothing in the output exists unless it's stated in the source, a closed world, every element must cite the exact source phrase in an evidence field, unknown values are null and ambiguous types are "ambiguous" rather than silently filled in, structural inferences are explicitly tagged "inferred": true with an "inference_basis", and elements that are mentioned but insufficiently defined get "status": "MISSING" with a "missing_reason" - it never introduces an equation or value absent from the source text. It classifies the problem as optimization, classification, simulation, proof, estimation, or other, then extracts five component types - Variables (id, name, symbol, type, domain, units, role, evidence), Operators (id, name, symbol, arity, inputs and outputs), Constraints (id, type, expression, variables involved, hardness), Objectives (id, direction - minimize, maximize, satisfy, find, or prove - expression, variables involved), and Uncertainty (type - stochastic, epistemic, measurement, model, or none stated - and what it affects) - each carrying its own evidence and inference or status flags. It surfaces what the document implies but doesn't state as a missing_information list, then validates and scores the whole extraction with flags for complete objectives, bounded variables, full evidence coverage, inference and missing counts, and an overall HIGH/MEDIUM/LOW formalizability rating. The result is a single JSON object with mps_version, source_title, problem_class, and the five component arrays plus missing_information and validation_flags.

Use it to formalize a problem statement into math, extract the mathematical structure from a research paper section, identify the variables, constraints, and objectives in a spec, convert a word problem into a structured MPS, or find what's missing from a problem formulation. It requires enough source text to cite every extracted element - a sparse prompt should come back with explicit missing-information markers rather than invented content - and it produces a formal specification only, not a solved optimization model or a proof.

Input: document text, a research excerpt, a problem description, or a specification. Output: a single MPS JSON object with problem_class, variables, operators, constraints, objectives, uncertainty, missing_information, and validation_flags including an overall formalizability rating.

• No external tools required for the core extraction - it's a structured output contract applied to the given text
• A full BYOK, bring-your-own-key, hosted version and the source repository (thebrierfox/doc2math-skill) are available separately, built by IntuiTek¹ under the MIT License

Researchers or engineers who need a rigorously grounded, evidence-cited mathematical formalization of a problem statement - including an honest account of what's missing - rather than a model that fills gaps with plausible-sounding assumptions.