AI4OrgChem

Release Validation License Evidence WSL2 AI assistance

中文说明

AI4OrgChem is an AI for Science (AI4S) Agent for independent computational reconstruction and evidence assessment of counter-traditional propositions in organic structure theory. The project was motivated by Zhong-Heng Yu's monograph, Questioning Fundamental Principles of Organic Chemistry (2024), and was implemented without using the monograph's program code.

AI-assisted research and engineering: OpenAI Codex (GPT-5.6). Project authorship, scientific decisions, interpretations, and publication responsibility remain with Xiao Chen; OpenAI is not presented as a project author, scientific certifier, peer reviewer, or institutional endorser.

Release status: Science Enhancement Release v0.2.0. The associated manuscript has not yet undergone peer review.

Science V0.2: the self-contained evidence package adds completed WP1 cyclobutadiene multireference evidence and WP3 benzene-mechanism intervention, an auditable WP2 backend-contract failure, an explicit unresolved WP4 outcome, and an internal clean evidence replay. The frozen V0.1 P01-P14 classifications are unchanged.

Post-V0.2 WP2 update: Science V0.3 WP2 incremental evidence closes the open-source cross-program reproducibility question across PySCF, Psi4, and NWChem: 8/8 frozen anchors and 2/2 relative-energy pairs pass. The historical ORCA-specific lane remains unexecuted, and no V0.1 proposition classification or V0.2 artifact is changed.

Textbook ideas such as conjugative stabilization, conjugation-driven planarization, steric destabilization, and aromatic stabilization are useful chemical heuristics. Problems arise when a heuristic is promoted to an unconditional mechanistic law. AI4OrgChem converts fourteen major propositions into falsifiable computational tasks with frozen systems, state definitions, sign conventions, numerical outputs, and explicit scope boundaries.

The project asks whether independently reconstructed quantum-chemical evidence supports, contradicts, or only partially supports each proposition. AI assists protocol management, evidence tracing, bounded molecular learning, active sampling, symbolic testing, and explanation; it does not assign scientific signs or replace electronic-structure calculations.

All fourteen propositions received determinate classifications within their declared tested domains:

• 12 consistent or scope-consistent with the corresponding monograph proposition;
• 2 partially consistent: P11 retains an opposite-sign inductive component, while P12 retains the boundary that onset sizes from different estimands are not directly comparable;
• 0 globally inconsistent and 0 unknown.

Representative frozen results are:

These results support a bounded methodological conclusion: several textbook heuristics do not automatically provide universally sufficient mechanistic explanations. They do not establish that traditional organic chemistry is globally wrong, do not create a universal opposite law of conjugative destabilization, and do not constitute institutional certification of the monograph.

The fourteen propositions form a layered, auditable evidence chain rather than fourteen equal-status quantum-chemistry reproductions. Independent QM, source-aligned/source-proxy reconstruction, published-value reanalysis, and graph/literature evidence retain distinct identities.

The 12+2 numbers are proposition-level classifications, not a count of fourteen mutually independent quantum-chemistry validations. P07 is a DERIVED synthesis of P04-P06 and adds no new QM evidence. P14 now publishes its lower-level records, source-proxy coordinates, and deterministic decision rebuild, while remaining limited to one C12H6 system.

See the P01-P14 evidence matrix and the evidence collection.

Quantum-chemistry experts who do not work with GitHub, AI, or software engineering can begin with the Rapid Review Guide for Quantum-Chemistry Experts: understand the overall argument in 10 minutes and audit any single proposition in 30 minutes without reading code first.

Independent reproduction, falsification, and scoped scientific disagreement are welcome. Use the repository's structured issue forms to submit an independent reproduction, scientific disagreement, documentation correction, or new molecular test proposal.

The completed bounded engineering line connects frozen scientific evidence to machine-readable data, equivariant learning, active-learning return, symbolic discovery, and a read-only evidence agent.

• bounded dataset: 17 geometries, 3 molecular families, and 5 energy targets;
• pi-pi family-holdout macro RMSE: 108.0 meV/atom for MACE and 108.2 meV/atom for NequIP;
• active learning: acquisition succeeded, while post-return model effects were mixed;
• PySR: the bounded pi-pi blind test passed, while the pi-sigma test failed;
• evidence agent: answers are restricted to frozen evidence and must expose sources and scope.

The dataset is too small for industrial or universal molecular generalization. Details are provided in the Agent capabilities and results, machine-readable evaluation summary, and limitations.

WSL 2不是这些量子化学公式成立的数学前提,但它是本项目实际形成并验证科学结果的权威运行平台。因此,"必须在WSL 2算"的准确含义是:若要声称与本项目运行环境等价,必须在已经验证的WSL 2 / Ubuntu 24.04软件栈中复核;原生Windows只用于阅读、轻量检查和发布整理,不被认证为PySCF科学主线、MACE/NequIP CUDA训练或PySR搜索的等价环境。

选择该平台的工程原因包括:Linux优先的PySCF与科学Python依赖;WSL GPU透传下已经跑通的CUDA/PyTorch链;MACE与NequIP不兼容e3nn版本的环境隔离;POSIX Shell入口;以及统一的线程、内存、进程和可写scratch边界。WSL环境是"经验证的平台合同",不是把Windows目录整体复制进Linux,也不允许覆盖用户已有环境。

本项目在WSL 2中执行了五类任务:1公开结构/光谱数据回归和数值账本重算;2约束或弛豫PES、几何优化与电子能/核排斥分解;3LFMO子空间、条件SCF、DSI/FUD/FUL/PDSI/GL/PLG状态及独立能量组装;4芳香/反芳香虚拟参考、LDE/ESE/VDE/ADE/CESE和图论规则检验;5第五期的MACE/NequIP训练、真实QM标签回流和PySR盲测。前四类形成P01–P14科学证据;第五类只检验AI4S工程可行性,不反向改变十四项判定。

并非十四项都是昂贵量子化学任务。P02主要是公开表值统计复核,P12是源账本与独立趋势比较,P13包含图论和分子式校验;表中明确列出每项在WSL 2上实际执行的任务。

统一机器定判入口为validate_public_evidence.py;双语导航和P11/P12差异由validate_evidence_navigation.py复核。定判入口读取冻结结果,不重新伪装成一次完整历史计算。

冻结命题与protocol
 → 环境/线程/内存预检
 → 分子、坐标、角度或公开数值账本
 → SCF/条件SCF/几何优化/PES/图论或统计任务
 → LFMO π/σ分类与状态构造(适用时)
 → 独立能量组装、收敛与物理诊断
 → processed JSON/JSONL冻结结果
 → 命题验证器与范围化判定
 → 数据卡、报告、文件清单和SHA-256快照

非QM命题跳过SCF步骤,但仍执行相同的协议冻结、来源追踪、机器结果和验证门禁。

• 先冻结分子、Cartesian几何或source-proxy、电子态、方法、基组、符号和完成规则,再运行;
• 同一比较必须保持同几何、同基组和同状态合同;不同Target或不同估计量不得直接相加;
• SCF收敛、电子数、自旋、占据、LFMO连续分类、内存和能量闭合失败必须保留为失败,不得静默删点;
• 历史坐标或程序不完整时必须标记source-proxy,不得宣称原厂代码或历史坐标身份复现;
• 默认8 CPU线程;中高内存QM任务串行;低内存门槛触发停止,不用swap掩盖OOM;
• 不覆盖现有micromamba环境;MACE与NequIP运行时分离;公开复核环境只在不存在时显式创建;
• AI预测、PySR公式和LLM解释不能生成或改写科学生产标签;P01–P14只由冻结量子化学/数值证据定判;
• 不使用原著程序代码,不把受版权保护全文、API密钥、本机路径、缓存或模型权重写入发布快照。

AI4OrgChem/
├─ project/ # 背景、研究项、价值、成果与十四项总表
├─ evidence/P01-P14/ # 每项data-card、protocol、result、report
├─ software/ # LFMO/条件SCF核心、64项测试、验证与刷新入口
├─ reproducibility/ # 双语运行手册、平台矩阵、环境与WSL入口
├─ configs/qm/ # 可公开的冻结量化配置
├─ manuscripts/ # Evidence matrix and publication positioning
├─ ai4s-agent/ # AI工程能力、评估、治理与限制
├─ figures/ # 项目自行生成的图
├─ manifests/ # 实质文件清单与SHA-256快照
└─ .github/workflows/ # 不调用Secret/LLM/昂贵QM的公开CI

原始量化输出、scratch、历史开发日志、模型权重、私有运行目录和原著文件不属于本精选归档。

wsl -d Ubuntu-24.04

进入仓库根目录后:

micromamba env list
# 仅当 ai4orgchem-public 不存在时:
micromamba create -f reproducibility/environment.yml
source reproducibility/wsl/activate-ai4orgchem-public.sh
bash reproducibility/wsl/ai4orgchem-verify
python software/scripts/validate_evidence_navigation.py
python software/scripts/validate_wsl_release.py
python software/scripts/validate_release_package.py
cd software
python -m pip install -e ".[science,test]"
python -m pytest -p no:cacheprovider

上述步骤复核冻结证据和公开核心软件,不承诺从精选包重跑所有历史昂贵作业。完整说明见中文运行手册。

结果文件不能手工改值。只有在协议授权的源数据、代码或文档发生变更并完成复算后,维护者才可在仓库根目录执行:

python software/scripts/refresh_release_snapshot.py
python software/scripts/validate_release_package.py
python software/scripts/validate_public_evidence.py
python software/scripts/validate_evidence_navigation.py
python software/scripts/validate_wsl_release.py

刷新只重建文件清单和SHA-256快照,不生成量子化学结论;随后必须重新运行64项测试、洁净克隆和Ubuntu入口验证,并创建新提交/不可变版本标签。外部复核者通常只需验证快照,无需刷新。

• P01–P14证据导航
• 十四项总证据矩阵
• 中文运行手册 / English runbook
• 中文平台矩阵 / Runtime platform matrix
• WSL双目录入口说明
• 公开核心软件与验证器
• AI4S Agent能力与结果及限制
• 文件清单与SHA-256 manifest

Validate the packaged P01-P14 result integrity without rerunning expensive quantum chemistry:

python software/scripts/validate_public_evidence.py
python software/scripts/validate_evidence_navigation.py
python software/scripts/validate_wsl_release.py
python software/scripts/validate_release_package.py

Expected output includes "status": "PASS", "propositions_checked": 14, and "propositions_navigated": 14.

Install and run the public core tests:

cd software
python -m pip install -e ".[science,test]"
python -m pytest -p no:cacheprovider

The canonical scientific runtime is WSL 2 / Ubuntu 24.04. Native Windows is supported only for limited evidence inspection and is not claimed to be equivalent to the validated PySCF, MACE/NequIP CUDA, or PySR runtime. See the runtime platform matrix, English runbook, and portable WSL entry points.

• The original monograph, scans, publisher files, full-text extracts, and historical program code are not distributed here.
• Some historical Cartesian coordinates and software were unavailable; affected results are explicitly marked as source-proxy rather than identity reproductions.
• Targets with different state contracts must not be summed across protocols.
• AI model outputs are engineering evidence, not new quantum-chemical labels or independent proof of the scientific propositions.
• Models, private run directories, caches, API credentials, and copyrighted source materials are excluded from this release.

• Evidence matrix
• Publication positioning
• 中文发布定位

Project-authored software and documentation are released under the Apache License 2.0. See NOTICE for attribution and third-party boundaries. The license does not relicense the monograph or other third-party material.

The sole project author is Xiao Chen. Contact: chenxiao0101@gmail.com. AI-assisted research and engineering support was provided through OpenAI Codex (GPT-5.6). Authorship, CRediT contributions, AI-assistance disclosure, and the competing-interests statement are recorded in AUTHORS.md. Machine-readable citation metadata is provided in CITATION.cff. Affiliation and ORCID are omitted because they were not supplied.