MathKernel Released to Resolve LLM Mathematical Errors with Evidence-Based Verification and MCP Server Support

To overcome the limitation that LLMs are strong in interpreting mathematical intent but vulnerable in precise arithmetic calculations, MathKernel has been released. This project adopts a division of labor structure where the LLM handles only parsing and planning, while a dedicated kernel performs actual calculations and evidence acquisition.

Core Architecture and Evidence Model

MathKernel is not a simple calculator but an Evidence-aware multi-engine orchestration layer. Every mathematical result includes a clear Trust Level and an Evidence Bundle as follows:

• Formal: Formal proofs verified by the Lean proof kernel
• Exact: Exact arithmetic operations or verified Certificates
• Symbolic: Consensus from symbolic engines such as SymPy
• Certified Numeric: Strict Interval Enclosure via tools like mpmath
• Empirical/Numeric: Empirical evidence or floating-point calculations

Agreement from a single engine is not considered proof; reliability is distinguished through independent cross-verification or formal verification. In particular, it tracks the Ancestry of Approximate Inputs and performs strict evidence management, such as automatically capping the trust level at 'Numeric' if decimal inputs are included.

Key Features and Tools

• Support for Diverse Mathematical Domains: Supports a wide range of areas including symbolic algebra, calculus, complex analysis, finite groups, linear algebra, graph theory, probability and statistics, and PDE/FEM (Finite Element Method).
• MCP (Model Context Protocol) Server: Through mathkernel-mcp, LLM agents can call 167 math tools. Agents explore capabilities via math_capabilities and request calculations and verification via math_reason.
• Multimodal Artifacts: Converts calculation results into portable HTML artifacts that can be visualized (Viz) and sonified, preserving the evidentiary reliability of the original data throughout the process.
• Performance Optimization: Ensures performance for large-scale computations such as Collatz conjecture spreads and GF(2^m) operations by leveraging Numba JIT compilation, CUDA GPU acceleration, and parallel processing pools.

Installation and Usage

It can be installed as a Python library (pip install mathkernel) and an MCP server (pip install 'mathkernel[mcp]'). Lean 4 and Mathlib are automatically installed upon first execution. This tool is useful for developers who want to reduce Hallucination when LLM agents handle math problems and clarify the source and verifiability of calculation results.