MathKernel: A Multi-Engine Mathematics Runtime Enhancing Trust and Provenance for LLMs

MathKernel is a new multi-engine mathematics kernel that enables applications and large language models (LLMs) to perform sophisticated mathematical computations with explicit evidence tracking. Available as a Python library and an MCP server, MathKernel separates the interpretation of mathematical intent, handled by the LLM, from the rigorous computation and evidence generation performed by the kernel. Each mathematical result includes a trust level, engine identifier, and a detailed derivation trail to ensure transparency and reliability. Unlike traditional systems where exact arithmetic is often conflated with formal proof, MathKernel distinguishes between exact computations, symbolic results, formal proofs, certified enclosures, and empirical evidence. This approach addresses the limitations of LLMs, which excel at understanding mathematical intent but are prone to errors in arithmetic. MathKernel delegates parsing, planning, and interpretation to the model, while the kernel executes computations and maintains claim-specific evidence. The system functions as a typed orchestration layer coordinating multiple mathematical engines, including symbolic, exact, certified, formal, and numerical solvers. It manages parsing, context handling, object identity, persistence, and evidence composition, while domain adapters perform the actual mathematical operations. Presentation layers are downstream and cannot alter the underlying mathematical claims. MathKernel supports a broad range of mathematical domains such as algebra, topology, probability, statistics, finite dynamics, PDEs, geometry, optimization, and stochastic differential equations. It provides typed objects for complex structures including distributions, time series, finite groups, manifolds, simplicial complexes, and more. The kernel rigorously tracks provenance and evidence, ensuring that approximate inputs do not silently degrade trust or proof strength. The MCP server interface exposes over 160 compositional operations and supports asynchronous job execution. MathKernel also offers standalone modules for specialized tasks like finite field arithmetic, Koopman analysis, and computational geometry. Performance optimizations include JIT compilation, GPU acceleration, and process isolation for external solvers, all while preserving the mathematical contract and evidence integrity. The system enforces strict separation between exact and numeric computations to maintain trustworthiness. Visualization and sonification tools transform MathKernel results into interactive, evidence-preserving artifacts. These tools produce portable HTML outputs that embed provenance metadata and maintain the trust model without upgrading evidence through presentation. MathKernel’s design emphasizes transparency, reproducibility, and rigorous evidence management, making it a valuable tool for research and applications requiring trustworthy mathematical computations integrated with LLMs. The project is open source under the MIT License and actively maintained with comprehensive testing and documentation.