academic-research-kernel

English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Deutsch · Français · Español · Português · Русский · Bahasa Indonesia · Italiano · हिन्दी · العربية · বাংলা · اردو · Tiếng Việt · Türkçe · فارسی · Kiswahili · Polski

A deterministic scholarly verification and manuscript engineering kernel for researchers and AI agents.

When researchers use large language models to read or write papers, two failure modes recur. While reading external papers, a model generates text probabilistically and readily repeats miscalculated $p$-values, shifted ablation tables, retracted references, or hollow code repositories as if they were verified facts. While writing your own manuscript, asking a model in prose to check formatting cannot reliably catch residual author metadata in double-blind PDFs, transcription errors in LaTeX tables, dangling figure references, or missing required venue sections.

Built from first principles, this repository separates semantic reading from deterministic verification: your language model handles reading and synthesis, while deterministic code handles mathematical recomputation, public database queries, syntax-tree checks, and SHA-256 provenance verification. The repository unifies 14 scholarly workflows, 18 standard Model Context Protocol (MCP) tools, and 16 CLI command groups on a single state kernel, working out of the box with Claude Code, Cursor, Codex, Gemini CLI, and Hermes Agent. Validation scope and external service limits are documented in the project audit.

Author: Junfu Shi (SJF, xngg1021). License: Source Lineage License 1.0.

1. Two-Minute Quick Start

Install uv, then choose either the local CLI + agent skill path or the stdio MCP server path for version 2.4.2:

# Option 1: Install the full CLI with digital PDF support and export the paper-research skill
uv tool install "academic-research-kernel[research]==2.4.2"
academic-research-kernel research skill --output ./paper-research

# Option 2: Start the stdio MCP server for Claude Code, Cursor, Codex, or Gemini CLI
uvx academic-research-kernel mcp

Run an offline environment check and try source normalization or the built-in venue specification catalog right away:

academic-research-kernel doctor
academic-research-kernel verify-source https://doi.org/10.1038/nature14539 --normalize-only
academic-research-kernel venue-template --action list-venues

The doctor command runs offline and checks runtime dependencies, all 26 JSON schemas, and all 18 MCP tool definitions; leaving optional external API keys unconfigured will not fail the check. Running mcp in a terminal waits silently on standard input/output for JSON-RPC requests from an MCP client. The 2.4.2 release page lists package SHA-256 hashes. Python 3.10–3.14 is supported, and statistical libraries load on demand only when distribution or power calculations are invoked.

2. First Half: Reading External Papers & Auditing Evidence

When reading external papers or conducting a literature review, the workflow moves from verifying bibliographic and repository authenticity, to recomputing reported statistics, to generating page-anchored reports.

1. Read a Paper or Compare Two Papers with Your Local Agent

Point your local agent at the exported ./paper-research/SKILL.md, pass a DOI, public URL, or local bu

... View full README on GitHub