MathModel-Agent

A closed-loop AI research agent framework for mathematical modeling competitions. 面向数学建模竞赛的闭环研究 Agent 框架：题目理解 → 选题 → 建模 → 求解 → 审查 → 复盘学习。

⚠️ 边界声明

• 这不是独立 Agent，需要配合 Executor/Critic/Controller 使用
• 这不是通用数学求解器，不替代 MATLAB/Python 计算
• 这不是论文生成器（Writer 需另配）
• 不保证获奖：竞赛有随机性，方法与执行质量才是可控变量

Overview

A modular agent framework that integrates:

• Problem parsing (Problem Graph + Assumptions + complexity score)
• Topic scoring for question selection
• Literature-driven architecture selection (architecture candidate pool)
• Baseline verification (baseline-gate, prevents over-optimization)
• Parallel solving (multi-seed, 32-core aware, checkpoint-resume)
• Automatic criticism (failure detection, proof trace, consistency, reviewer, compliance)
• Structured math scaffolding (per-problem-type math reasoning scaffold with L1/L2 tiers)
• Memory replay (failure/success/strategy/decision libraries, experience evolution)

Architecture

Problem
  ↓
Problem Parser (Problem Graph + Assumptions + complexity)
  ↓
Topic Scorer (experience coverage → question score)
  ↓
Literature → Architecture Candidates → Baseline Gate
  ↓
Solving (parallel, checkpoint, result-only-from-code)
  ↓
Critic (failure → proof → consistency → reviewer → compliance)
  ↓
Delivery Gate (visual contract + math structure checks)
  ↓
Memory (replay, experience evolution)

Key Features

1. Closed-loop Research

Failure → Diagnosis → Memory → Replay → Improvement

2. Three Permission Modes

Mode	When	Scope
AUTO	Autonomous research	Low-risk decisions self-decided, with emergency stop
COLLAB	Human in loop	High-impact decisions need approval
LOCK	Submission freeze	Frozen: suggest only, no changes

3. Math Reasoning Scaffold（数学推理脚手架）

按题型注入"数学化骨架"（符号化 → 目标 → 约束 → 求解框架 → 理论性质）， 分 L1（轻量，默认）/ L2（完整，含反锚自检）两层，防止建模过程退化为"模型调用说明书"。

4. Experience-driven Memory

The skill is an experience-driven agent:

• 多类型经验库（failure/success/strategy/decision）+ 写作经验库
• 每条经验带质量评分（source_quality / confidence / verified / reuse_count / success_rate）
• 检索：字符指纹 → n-gram TF-IDF →（可选）embedding 混合召回 + 分档排序
• 赛后复盘 → 经验更新 → 越比越强

Quick Start

安装：将本仓库内容放入 Agent 的 skills 目录 （如 ~/.claude/skills/contest-orchestrator/ 或对应技能的 skills 目录）。

开始建模，我选A题

Pipeline: parser → planner → solver → critic → report.

Modules

阶段	入口
权限/模式	core/permission.md
路由/止损	core/controller.md
时间/冻结	core/time_engine.md
选题/架构/模型	planner/problem_parser.md, planner/planning.md
求解/出图/写作	executor/execution.md
视觉排版	output_contract.md, presentation_polish.md
检查/证明/合规	critic/criticism.md
团队协议	team/interface.md

Repository Layout

core/        brain rules (permission, controller, time_engine, delivery gate)
planner/     think (problem_parser, planning)
executor/    act (execution)
critic/      review (criticism)
scripts/     validation and routing checks
team/        collaboration protocol

Contributing

欢迎贡献竞赛经验（复盘、失败/成功案例）——不只代码。

License

Apache-2.0（见 LICENSE）。