测试版 v0.1.0 · 作者已用于日常研究，尚无外部用户完成测试 · 报告问题Beta v0.1.0 · used daily by its author, but not yet tested by other users · Report an issue

Framework for Unified Scientific Intelligence in Open Nuclear physics

核物理程序和文献，
放在一个终端里 Nuclear codes and papers
in one terminal

FUSION 的核心是三层知识库：一份随仓库发布的离线文献库，arXiv nucl-th 的 61,059 篇论文每篇一页，加 108 个主题页和一张引用图；一份你读过的文献笔记；一份你自己的研究档案。 agent 在这三层上查文献、记结论、避开走过的弯路，再用 20 个程序技能去跑具体计算。 全部是纯文本，不需要服务器或数据库；DeepSeek、Qwen、GLM、Claude 和 GPT 都能接，多步任务上差距明显。 背景中的每个光点来自语料库里的一篇论文。 At the core of FUSION is a three-layer knowledge base: an offline literature corpus that ships with the repository, one page for each of 61,059 arXiv nucl-th papers plus 108 topic pages and a citation graph; your own notes on the papers you have read; and your own research record. The agent works across the three to find literature, keep conclusions, and avoid dead ends already walked, then runs calculations through 20 code skills. Everything is plain text, with no server or database. It runs with DeepSeek, Qwen, GLM, Claude, and GPT, though not equally well on multi-step work. Each point of light in the background represents one paper in the corpus.

了解三层知识库See the knowledge base → 26 个技能26 skills 查看源码View source
FUSION logo
61,059arXiv nucl-th 论文页arXiv nucl-th paper pages
61,167离线 Markdown 知识库页面offline Markdown knowledge-base pages
809,632语料内引用边in-corpus citation edges
26技能随仓库发布，其中 20 个负责具体程序shipped skills, including 20 for specific codes

是什么 · WhatWhat it is

FUSION 分成三层 One platform, three layers

底层是 opencode 的品牌分叉（MIT，每周由 CI 重基上游），中间放核物理技能和知识库，最上层挂载用户自己的文献笔记、研究档案与密钥。 The base is a rebrand fork of opencode (MIT, rebased onto upstream by weekly CI). Nuclear-physics skills and the corpus form the middle layer. Users mount their own notes, research profiles, and credentials on top.

FFramework: 技能和知识库可以继续增加skills and corpus modules can be added
UUnified: 程序、文献和算力共用一个入口one interface for codes, literature, and compute
SScientific: 每项能力都用可复核的数值检验each capability is checked against numerical references
IIntelligence: agent 负责安装程序、编写输入、运行计算并解析结果the agent installs, writes inputs, runs, and parses
OOpen: 只接入公开可获得、可构建的程序，并注明许可限制codes must be public and buildable, with licence limits stated
NNuclear physics: 覆盖反应、结构、核天体与重离子输运reactions, structure, nuclear astrophysics, and heavy-ion transport
私人层 · 不进入公开发行版Private layer · never shipped

你的文献 wiki、研究档案、密钥Your literature wiki, research profile, credentials

FUSION 只定义挂载接口，私人数据始终留在用户自己的机器上。FUSION defines the mount interface; private data stay on the user's machine.

核物理定制层Nuclear-physics layer

程序技能 + 领域知识库 + 科研工作流Code skills + domain knowledge base + research workflows

仓库内共有 26 个技能：20 个负责具体程序，SFRESCO 用于拟合，EXFOR 获取实验数据，kb-search 检索离线知识库，literature-wiki 和 research-profile 维护个人研究 wiki，fusion-setup 负责首次配置。知识库的 61,167 个页面都能离线浏览。The repository ships 26 skills: 20 for specific codes, plus SFRESCO fitting, EXFOR data retrieval, offline knowledge-base search, two personal research-wiki tools, and first-run setup. All 61,167 knowledge-base pages can be browsed offline.

引擎 · opencode 品牌分叉Engine · opencode rebrand fork

只改品牌，不碰功能代码Brand patch only, functional code untouched

CI 每周自动重基上游，可接入 DeepSeek、Qwen、GLM、Claude 和 GPT。Weekly CI rebases the fork onto upstream. DeepSeek, Qwen, GLM, Claude, and GPT are supported.

↑ 上层依赖下层，越往上越私密each layer builds on the one below; privacy increases upward

知识库 · KnowledgeThe knowledge base

三层知识库：整个领域、你读过的、你做过的 Three layers of knowledge: the field, what you read, what you did

FUSION 的核心是三层 wiki。它们全是纯文本 Markdown，agent 用 grep 就能查，不需要服务器、数据库或向量索引。第一层随仓库公开发布；后两层属于你自己，只存在你的机器上。回答一个问题时，agent 依次查这三层：领域里谁做过，你读的时候得出了什么结论，你自己是不是已经试过、为什么没成。 The core of FUSION is three wikis. All three are plain Markdown that the agent searches with grep, with no server, database, or vector index. The first ships publicly with the repository; the other two are yours and stay on your machine. To answer a question the agent works down the layers: who in the field has done this, what you concluded when you read them, and whether you have already tried it yourself and why it did not work.

61,059论文页paper pages
108PhySH 主题页PhySH topic pages
809,632语料内引用边in-corpus citation edges
237,009标了类型的引用关系typed citation relations
kb-wiki/papers/2101.03193.md真实页面节选real page, excerpt
---
arxiv: 2101.03193
title: "Implications of PREX-II on the equation of state of neutron-rich matter"
authors: "Reed, Brendan T.; Fattoyev, F. J.; Horowitz, C. J.; Piekarewicz, J."
doi: 10.1103/PhysRevLett.126.172503
digest_model: deepseek-chat
digest_date: 2026-07-14
concepts: ["neutron-skin-thickness", "symmetry-energy"]
---
Laboratory experiments sensitive to the equation of state of neutron rich matter in the vicinity of nuclear saturation density provide the first rung in a "density ladder" … [arXiv abstract, verbatim]
## Key claim
Using a specific class of relativistic energy density functionals, the authors report a value of the slope of the symmetry energy L = (106 ± 37) MeV derived from the PREX-II neutron skin measurement …

## Key numbers
- Neutron skin thickness of 208Pb: Rskin = (0.283 ± 0.071) fm
- Slope of symmetry energy at saturation density: L = (106 ± 37) MeV
…
## Related work
contrasts1311.0168  Neutron skin of 208Pb from Coherent Pion Photoproduction
contrasts1904.12269  Neutron skins of atomic nuclei: per aspera ad astra
compares1611.01871  Symmetry Energy III: Isovector Skins
uses1306.6034   Has a thick neutron skin in 208Pb been ruled out?
uses1912.05702  A NICER View of PSR J0030+0451 …
…
## In-corpus citations
Cites (33)   Cited by (298)

引用关系按用途分类Citations typed by what they do

uses191,789
compares23,183
extends12,366
contrasts5,640
applies4,031

摘要逐字取自 arXiv；Key claim、Method、Key numbers 由语言模型读全文生成，不是作者写的，每页都记下了生成模型和日期。引用关系只有两个来源：论文正文里写明的 arXiv 编号和 DOI，以及 INSPIRE 的参考文献列表，不靠引用键去猜。它们是找文献的线索，写论文时仍要引用原文并回原文核对。作者要求撤下自己的页面，一律照办。Abstracts are verbatim from arXiv; Key claim, Method, and Key numbers are generated by a language model reading the full text, not written by the authors, and every page records the model and date. Citation links come from two sources only, arXiv ids and DOIs written in the paper itself and its INSPIRE reference list, never from guessing at citation keys. They are leads for finding literature: cite the paper, and check the link against it. Any author who asks for their page to be removed gets it removed.

~/research-wiki/目录结构layout
├── raw/            # 原始 PDF，只进不改# source PDFs, never edited
├── sources/        # 每篇读过的论文一页# one page per paper you read
├── methods/        # 壳模型、R 矩阵、耦合道 … 每个方法一页# shell model, R-matrix, CDCC … one page each
├── systems/        # ²⁰⁸Pb、⁴⁸Ca、¹²C … 每个体系一页# ²⁰⁸Pb, ⁴⁸Ca, ¹²C … one page per system
├── observables/    # 截面、谱因子 … 每个物理量一页# cross sections, widths … one page each
├── entities/       # 作者、组、实验、程序# authors, groups, experiments, codes
├── debates/        # 谁和谁的结论不一致# who disagrees with whom
├── synthesis/      # 跨论文的综合判断# conclusions that span papers
└── log.md
sources/<paper-id>.md页面骨架page skeleton
---
methods: [ … ]   systems: [ … ]   observables: [ … ]
source_pdf: raw/…
---
## For future agents
两三句话：论文主张什么、那个关键数字是多少、涉及哪些方法和体系、
什么问题会用到这一页。agent 读这几行就能决定要不要往下看。Two or three sentences: the claim, its one key number, the methods and
systems it touches, and which questions need this page. The agent
decides from these lines whether to read on.

## Key claims    带章节出处with section references
## Key numbers
→ [[methods/…]]  [[systems/…]]  [[debates/…]]

每收进一篇，通常会改动 5 到 15 个页面：方法页、体系页的正文被改写，新旧结论矛盾的地方被标出来，写进 debates/。标签来自受控词表，所以“用过 CDCC 的所有论文”这种问题一次就能答。Filing one paper typically touches 5 to 15 pages: method and system pages are rewritten, and conflicts with earlier conclusions are flagged in debates/. Tags come from a controlled vocabulary, so a question like “every paper I read that uses CDCC” is answered in one read.

~/research-wiki-personal/目录结构layout
├── profile.md      # 一页摘要，每次会话自动载入# one-page summary, loaded every session
├── projects/       # active / paused / done
├── papers/         # published / submitted / drafts / rejected
├── ideas/          # promising / parked / killed
├── failures/       # 没做成的事，以及原因# what did not work, and why
├── collaborators/
└── log.md
failures/<slug>.md页面骨架page skeleton
## What was tried
## Why it should have worked      # 当时的判断# the belief at the time
## What actually happened
## What was ruled out
## Conditions that might revive it

科研里最常见的浪费，是把同一条死路再走一遍。失败页和论文、项目用同一套标签（方法、体系、课题），所以 agent 在你开始一个新想法之前，能先查到自己以前在同一个体系上栽过什么跟头。状态变化（投稿、接收、换单位）按时间追加记录，不覆盖旧值。The commonest waste in research is walking the same dead end twice. Failure pages carry the same tags as papers and projects (method, system, topic), so before you start a new idea the agent can find what already went wrong on that system. Changes of state (submitted, accepted, a new affiliation) are appended with dates, never overwritten.

为什么是 wiki，而不是每次让模型现查Why a wiki, instead of asking the model fresh each time

读一次，一直有效Read once, useful forever

直接问模型，或者临时检索 PDF，每个问题都要从原文重新找一遍，读过的东西留不下来。wiki 把一篇论文读一次、整理一次，结论写进方法页和体系页，下一个问题直接在上面接着用。Asking a model, or retrieving from PDFs on the fly, rediscovers everything for every question, and nothing read is kept. A wiki reads and files a paper once, writes its conclusions into method and system pages, and the next question starts from there.

模型记不住，文件记得住The model forgets; files do not

每次开新会话，模型都从零开始。个人档案的一页摘要在会话开始时自动载入，模型一上来就知道你在做什么、试过什么、为什么放弃，不用你每次重讲一遍。Every new session starts from zero. The one-page summary of your research record loads at the start, so the model already knows what you are working on, what you tried, and why you stopped, without you retelling it.

先读摘要，再决定读不读Triage before reading

上下文窗口是有限的。每页开头几行就说清楚这篇讲什么、关键数字是多少，agent 扫过几十个候选页就能筛掉大半，把窗口留给真正相关的内容，而不是塞满整篇 PDF。The context window is finite. The first lines of each page state the claim and the key number, so the agent can scan dozens of candidates, drop most of them, and spend the window on what matters instead of whole PDFs.

每句话都能追到出处Every answer has a source

结论指向具体的页面，页面指向 arXiv 原文。检索规则要求报告的每一条结果都来自实际跑过的搜索，模型不能凭记忆编文献。页面是纯文本，人可以直接打开、改错，用 git 看改动历史。Conclusions point to pages and pages point to arXiv. The search rules require every reported hit to come from a search that actually ran, so the model cannot cite from memory. Pages are plain text: a person can open them, fix them, and read the history in git.

换模型，不用重建Swap models, keep the knowledge

没有向量数据库，也不绑定某个嵌入模型。DeepSeek、Qwen、GLM、Claude、GPT 读的是同一批文件，换模型或者断网都不影响。No vector database and no tie to an embedding model. DeepSeek, Qwen, GLM, Claude, and GPT all read the same files, so switching models or working offline changes nothing.

私人的留在本机Private stays local

公开发布的只有第一层。你读过什么、怎么评价、哪些想法失败了，都只在你自己的机器上，FUSION 只定义怎么挂载，不打包、不上传。Only the first layer is published. What you read, what you thought of it, and which ideas failed stay on your machine; FUSION defines how they mount and never packages or uploads them.

一个问题ONE QUESTION“²⁰⁸Pb 的中子皮到底有多厚？”“How thick is the neutron skin of ²⁰⁸Pb?”

1 · CORPUS

领域里谁说过什么What the field has said

grep 摘要和 Key numbers 找到 PREX-II 给出的 0.283 fm，再沿 contrasts 关系找到结论相反的测量（例如相干 π 光生），主题页给出全貌。Grep abstracts and key numbers to find PREX-II's 0.283 fm, follow the contrasts relations to measurements that disagree (coherent pion photoproduction, for one), and read the topic page for the landscape.

2 · LITERATURE WIKI

你读的时候怎么判断What you concluded

打开你读过的那几篇的页面和 debates/ 里的记录：哪些结论你当时就不信，哪些数字后来被别的论文推翻。Open your pages on the papers you read and the notes in debates/: which conclusions you doubted at the time, which numbers were later overturned.

3 · PERSONAL WIKI

你自己做到了哪一步Where your own work stands

查你的项目和失败页：这个体系你算过没有，卡在哪里，哪条路已经证明走不通。Check your projects and failure pages: whether you have computed this system, where it stalled, which routes are already known to fail.

实际使用 · In actionIn action

一次真实的对话 A real session

下面是一次录制下来的会话，Claude Code 配 Opus 5：一句自然语言，它取 KD02 参数、写输入卡、跑 FRESCO、做收敛检查，再去 EXFOR 查有没有对应能量的数据。命令和数字逐个照抄，只压缩了长度。 A recorded session, Claude Code on Opus 5: from one sentence it pulls the KD02 parameters, writes the deck, runs FRESCO, checks convergence, and asks EXFOR whether the energy was ever measured. Commands and numbers are copied from that recording; only the length is compressed.

fusion · claude code · opus 5
❯ 算 50 MeV 的 n+90Zr 弹性散射，用 KD02 全局光学势，然后跟 EXFOR 上有的实验数据比一下compute n+90Zr elastic scattering at 50 MeV with the KD02 global optical potential, then compare against whatever EXFOR has
◆ skill fresco loaded · install verified · benchmark current
◆ omp.py --code kd02 --proj n --target 90Zr --energy 50 → ap=0 (R = r₀At1/3)
✓ deck written → n90zr_50.in (potential terms: V, W_d, V_so + Coulomb)
✓ σ_R = 1301.638 mb → 1301.640 mb （步长减半、分波加倍、去掉截断）(step halved, partial waves doubled, truncation dropped) Δ 1.8e-6
✓ EXFOR 13160 · n+90Zr elastic → 8.0 10.0 24.0 MeV，没有 50 MeV, no 50 MeV
❯

为什么 · WhyWhy FUSION

通用 agent 容易漏掉这些核物理细节 Generic agents miss code-specific details

FRESCO 的 namelist 怎么写，TALYS 为什么报错仍返回 0，1990 年代的 nucl-th 预印本去哪里找，这些都需要专门知识。FUSION 把答案写进技能和离线语料库。 FRESCO namelists, unreliable TALYS exit codes, and nucl-th preprints from the 1990s all need code-specific handling. FUSION records that knowledge in skills and an offline corpus.

SKILLS

每个程序，一个专家技能One expert skill per code

20 个程序技能已经覆盖反应、裂变统计、R 矩阵与核天体、结构与从头算、TDHF、重离子输运和状态方程。每个技能都包含安装、输入、运行和结果解析，并且必须写明验证等级，复现论文结果或程序自带的参考值。连“TALYS 出错也返回 0”这种坑都写在里面。 Twenty code skills cover reactions, fission and statistical decay, R-matrix and nuclear astrophysics, structure and ab initio theory, TDHF, heavy-ion transport, and equations of state. Each teaches installation, input authoring, execution, and parsing, and must declare its validation tier and reproduce a published or code-shipped reference. Even traps such as “TALYS exits 0 on fatal errors” are encoded.

KNOWLEDGE

离线检索 nucl-th 语料Offline search across the nucl-th corpus

离线知识库共有 61,167 页，其中 61,059 个论文页收录元数据、arXiv 摘要、机器生成的内容摘要和引用链接，另有 108 个主题页。agent 用 grep 就能查，不需要服务、API key 或网络。这些页面可能有错，写论文时仍要引用原文。 The offline knowledge base contains 61,167 pages: 61,059 paper pages with metadata, arXiv abstracts, generated digests, and citation links, plus 108 topic pages. The agent searches them with grep, with no service, API key, or network required. These generated pages can be wrong; cite the original paper.

MODELS

支持多种模型Runs with several model providers

DeepSeek、Qwen、GLM、Claude 和 GPT 都可以用，Phase 0 的验收全程跑在 deepseek-chat 上，没有调用国外 API。但兼容不等于效果相同：单步任务国产模型没问题，而写输入卡、跑程序、读回结果再改一轮这类多步任务，deepseek-chat 只是概率性地做对。要稳定复现，目前还是 Claude Code 或 Codex 配各自最强的模型。 DeepSeek, Qwen, GLM, Claude, and GPT are supported, and the Phase 0 quality gate ran entirely on deepseek-chat, so it did not depend on access to an overseas API. Compatibility is not equivalence, though: single-step work is fine on any of them, but authoring a deck, running the code, reading the result back and revising it is a chain where deepseek-chat is right only some of the time. Reproducing that reliably still means Claude Code or Codex on their strongest models.

PRIVACY

私人数据留在本机Private data stay on your machine

精读文献 wiki、研究档案和集群密钥都挂载在私人层。FUSION 只定义接口，不会打包或分发这些文件。 Your read-literature wiki, research profile, and cluster credentials mount into the private layer. FUSION defines the interface and never ships anyone's personal data. The platform is open; your knowledge is private.

物理 · PhysicsThe physics inside

二十个程序背后的原子核物理 The nuclear physics behind the twenty codes

每张图对应一类技能，选的都是基准表里真正跑过的体系或同类的标准例子。图是示意，但账是真的：原子核按实际的质子数、中子数堆成，半径取 1.2A1/3 fm；裂变和俘获前后质子数、质量数守恒；壳层图的能级取自实测分离能。鼠标移上去可以转动视角。 Each panel stands for one family of skills, using the systems in the benchmark table or standard examples of the same kind. The pictures are schematic but the bookkeeping is real: nuclei are built from their actual proton and neutron numbers at radius 1.2A1/3 fm, fission and capture conserve Z and A, and the shell diagram uses measured separation energies. Move the pointer over a panel to tilt it.

FRESCOCOLOSSNLATCNOKpikoeSIDESSWANLOP

直接反应：氘核在 ²⁰⁸Pb 表面破裂Direct reactions: a deuteron breaks up on ²⁰⁸Pb

弱束缚的氘核（结合能只有 2.22 MeV）掠过铅核表面时被拆开：中子被靶核吸收，质子在库仑场里偏转飞走，这就是非弹性破裂。靶核由 82 个质子和 126 个中子堆成，半径 1.2A1/3 ≈ 7.1 fm。The weakly bound deuteron (binding energy only 2.22 MeV) is torn apart as it grazes the lead surface: the neutron is absorbed and the proton is deflected by the Coulomb field, which is nonelastic breakup. The target holds 82 protons and 126 neutrons at radius 1.2A1/3 ≈ 7.1 fm.

反应类的 8 个技能覆盖光学模型、耦合道与 CDCC、转移和敲出反应；基准表里 FRESCO 与 COLOSS 两个独立求解器的反应截面一致到 6 位。The eight reaction skills cover the optical model, coupled channels and CDCC, transfer and knockout; in the benchmark table FRESCO and COLOSS, two independent solvers, agree on the reaction cross section to six figures.

CCFULLFRESCO

熔合：越过库仑势垒Fusion over the Coulomb barrier

¹⁶O 打 ¹⁴⁴Sm，两核接触后粘在一起，形成高激发的复合核 ¹⁶⁰Yb*。势垒附近的熔合截面对靶核的集体激发很敏感，这正是耦合道计算要处理的。¹⁶O on ¹⁴⁴Sm: once the surfaces touch, the two nuclei stick and relax into the excited compound nucleus ¹⁶⁰Yb*. Near the barrier the fusion cross section is sensitive to collective excitations of the target, which is what coupled-channels codes are for.

CGMFTALYS

裂变：²⁵²Cf 一分为二Fission: ²⁵²Cf splits

核被拉长、出现颈部后断开。这里画的是其中一种分法：初级碎片 ¹⁰⁸Mo + ¹⁴⁴Ba，各蒸发两个中子成为 ¹⁰⁶Mo + ¹⁴²Ba，再放出 γ 光子。²⁵²Cf 自发裂变平均放出约 3.8 个瞬发中子，CGMF 逐事件模拟这一步。The nucleus elongates, forms a neck and snaps. One of many splits is drawn: primary fragments ¹⁰⁸Mo + ¹⁴⁴Ba, each evaporating two neutrons to become ¹⁰⁶Mo + ¹⁴²Ba, then emitting γ rays. Spontaneous fission of ²⁵²Cf releases about 3.8 prompt neutrons on average; CGMF follows this event by event.

AZURE2SkyNet

核天体：¹⁶O(p,γ)¹⁷FNuclear astrophysics: ¹⁶O(p,γ)¹⁷F

恒星里，质子被 ¹⁶O 俘获成为 ¹⁷F，多出的能量由 γ 光子带走，这是 CNO 循环的一个分支。基准表里 AZURE2 复现的正是这个反应的天体物理 S 因子；SkyNet 把成千上万个这样的反应接成核合成网络。In a star, ¹⁶O captures a proton to become ¹⁷F and a γ ray carries off the excess energy, a branch of the CNO cycle. AZURE2 reproduces this reaction's astrophysical S factor in the benchmark table; SkyNet links thousands of such reactions into a nucleosynthesis network.

KSHELLNuclearToolkit.jlGSMSky3D

壳结构：¹⁶O 的双幻数Shell structure: doubly magic ¹⁶O

质子（左）和中子（右）各 8 个，恰好填满 0s½、0p3/2、0p½；再往上的 0d5/2 要多花约 11 MeV，这就是幻数 8。能级取自 ¹⁶O、¹⁵N/¹⁵O 和 ¹⁷O/¹⁷F 的实测分离能与激发能，0s 只示意位置；质子能级整体偏高，是库仑排斥的结果。Eight protons (left) and eight neutrons (right) exactly fill 0s½, 0p3/2 and 0p½; the next orbital, 0d5/2, costs about 11 MeV more, which is magic number 8. Levels come from measured separation and excitation energies of ¹⁶O, ¹⁵N/¹⁵O and ¹⁷O/¹⁷F, with 0s placed schematically; the proton levels sit higher because of Coulomb repulsion.

SMASHGiBUU

重离子碰撞：Au + AuHeavy-ion collision: Au + Au

接近光速的两个金核在实验室系被洛伦兹收缩成薄饼。重叠区的核子（参与者）停下来，变成高温高密的火球并产生大量强子；不重叠的旁观者继续往前飞。输运程序逐个粒子追踪这个过程。Two gold nuclei near the speed of light are Lorentz-contracted into pancakes. Nucleons in the overlap (participants) stop and turn into a hot, dense fireball that produces many hadrons, while the spectators fly on. Transport codes follow this particle by particle.

vHLLEThermal-FIST

流体与状态方程：椭圆流Hydrodynamics and the EOS: elliptic flow

非对心碰撞留下的火球是杏仁形的。反应平面内的压强梯度更大，所以沿短轴膨胀得更快，形状最后反过来；这种各向异性叫椭圆流，对状态方程和粘滞很敏感。颜色表示温度，冷却到冻结后强子自由飞出。An off-centre collision leaves an almond-shaped fireball. The pressure gradient is steeper in the reaction plane, so it expands faster along the short axis until the shape inverts; this elliptic flow is sensitive to the equation of state and viscosity. Colour tracks temperature; after freeze-out the hadrons stream freely.

实测基准 · BenchmarksBenchmarks already run

每个技能都要先跑通参考算例 Every skill must reproduce a reference result

当前 20 个程序技能覆盖反应、结构、裂变统计、核天体、重离子输运和状态方程。下表只写已经构建并跑出的结果，计划中的功能不算。 The 20 current code skills cover reactions, structure, fission and statistical decay, nuclear astrophysics, heavy-ion transport, and equations of state. The table reports completed builds and runs, not planned features.

8反应Reactions
FRESCOCCFULL COLOSSpikoe NLATCNOK SIDESSWANLOP
2裂变与统计衰变Fission and statistical decay
TALYSCGMF
2R 矩阵与核天体R-matrix and nuclear astro
AZURE2SkyNet
4结构、从头算与 TDHFStructure, ab initio, and TDHF
GSMKSHELL NuclearToolkit.jlSky3D
4重离子输运与状态方程Heavy-ion transport and EOS
SMASHGiBUU Thermal-FISTvHLLE
程序Code 基准用例Benchmark case 复现精度Agreement
TALYS T1 发行版自带 1,438 个参考输出文件the 1,438 reference output files shipped with the distribution 1,419 个逐字节一致，其余 18 个数据文件约 6 位有效数字1,419 byte-for-byte; the remaining 18 data files to ~6 sig figs
NuclearToolkit.jl T1 完整 30 项测试：手征 EFT、HFMBPT、IMSRG、壳模型full 30-test chain: chiral EFT, HFMBPT, IMSRG, and shell model 30/30；⁴He 基态能到 10⁻⁶30/30; ⁴He ground-state energy to 10⁻⁶
Sky3D T1 ¹⁶O 静态 TDHF：3,268 个能量、单粒子与多极矩量¹⁶O static TDHF: 3,268 energy, single-particle, and moment values macOS/Linux 的输出在打印精度内全部一致，均迭代 370 步all printed values match on macOS/Linux, with the same 370 iterations
SkyNet T1 核合成网络自带比较套件与 NSE 物理锚点nucleosynthesis network comparison suite plus an NSE physics anchor Linux 19/19；macOS 17/19，两个平台差异明确标注19/19 on Linux; 17/19 on macOS, with both platform limits stated
CGMF T1 ²⁵²Cf 自发裂变与热中子诱发 ²³⁵U 的 40 事件历史40-event histories for ²⁵²Cf spontaneous fission and thermal-neutron-induced ²³⁵U 与 LANL 发行版参考逐字节一致byte-for-byte identical to the LANL-shipped references
SMASH T1 104 项测试与 Au+Au 输运守恒律104-test suite and conservation laws in Au+Au transport Linux 104/104；两平台 B = 788、Q = 316 精确守恒104/104 on Linux; exact B = 788 and Q = 316 on both platforms
AZURE2 T2 ¹⁶O(p,γ)¹⁷F，按论文表格重建 9 个参数，不做拟合¹⁶O(p,γ)¹⁷F, nine parameters reconstructed from the paper, with no fit S(90 keV) 偏差 -5.7%；对实测数据 χ²/N = 1.53S(90 keV) differs by -5.7%; χ²/N = 1.53 against measured data
FRESCO × COLOSS n+⁹⁰Zr 弹性散射 50 MeV，KD02 全局光学势；两个独立求解器互检（Numerov 耦合道 vs 复标度 Lagrange-Laguerre）n+⁹⁰Zr elastic at 50 MeV, KD02 global OMP; two independent solvers cross-checked (Numerov coupled-channels vs complex-scaled Lagrange-Laguerre) σR 一致到 6 位；步长收敛 9 位σR agrees to 6 sig figs; converged to 9
最新发现LATEST FINDING

使用相同的源码、输入和随机种子，SMASH 在 macOS 与 Linux 上得到的部分粒子多重性最多相差 25%，但两边的重子数 788 和电荷 316 都精确守恒。验证器据此检查守恒律，不再把依赖平台的多重性当作“标准答案”。这次测试还发现，旧规则会把轻核的 PDG 编码误判成重子数 0。 With identical source, input, and seed, some SMASH multiplicities differ by up to 25% between macOS and Linux, while baryon number 788 and charge 316 remain exact integers on both. The verifier therefore anchors on conservation laws instead of treating platform-sensitive multiplicities as ground truth. The same check caught an old rule that assigned baryon number zero to light-nucleus PDG codes.

T1 要求复现程序随发行版提供的参考输出或测试套件。没有这类参考值时使用 T2，改用跨平台构建、物理恒等式和实验数据检验。每个技能都会写明等级、适用范围和已知失败模式，并接受对抗检查和引文核验。 T1 reproduces references or a test suite shipped by the code itself. T2 is used when no distributable reference exists, so cross-platform builds, physics identities, and measured data form the evidence chain instead. Every skill declares its tier, scope, and failure modes before adversarial review and citation verification.

真实任务 · Case studiesTwo real runs

两个跑完的任务，有结果，也有空缺 Two complete tasks, including what they did not find

这两个任务都实际跑过，不是预先写好的演示。第二个任务没有找到目标能量下的实验数据，页面也如实写明了这个空缺。 These are completed research tasks, not scripted demos. The second task found no measurement at the requested energy, and reports that absence as part of the result.

CORPUS

从一篇 PDF 找到相关工作Finding related work from one PDF

输入只有一篇 PDF（Abu-Ibrahim 等，PRC 77, 034607，碳同位素在质子靶上的反应截面）。整个任务离线完成，没有调用外部 API。 The input was a single PDF (Abu-Ibrahim et al., PRC 77, 034607, reaction cross sections of carbon isotopes on a proton target). Fully offline, zero external API calls.

语料库先把论文定位到 0710.4193。引用网络找到它引用的 8 篇论文和引用它的 5 篇论文，其中包括同组前作 nucl-th/0612029、²²C 双中子晕工作 nucl-th/0605055和黑球近似 nucl-th/0410032。词法检索又补出方法相近、研究体系不同的工作，如氧同位素 Glauber 计算和相对论碰撞近似。现成摘要里还有可直接核对的数字：p+¹²C 40 MeV 的 σR = 432 mb，²²C 的 rm = 3.6 fm，经验公式 R(C) = 0.96 ± 0.05 覆盖 153 个数据点。 The corpus resolved it to 0710.4193; the citation graph returned the 8 papers it cites and the 5 citing it (including the group's own predecessor nucl-th/0612029, the ²²C two-neutron halo nucl-th/0605055, and the black-sphere approximation nucl-th/0410032), and lexical neighbours added same-method, different-system work. The pre-generated digest supplied checkable numbers: σR = 432 mb for p+¹²C at 40 MeV, rm = 3.6 fm for ²²C, and the empirical R(C) = 0.96 ± 0.05 across 153 data points.

限制：引用关系只在这批语料内统计。RIKEN 的实验论文不属于 nucl-th，所以“被引 5 次”低估了实际引用量。完整引用数仍要到 INSPIRE 查询。 Limit: citation edges are counted only inside this corpus. The RIKEN measurements are not nucl-th papers, so “cited by 5” understates the true count. An authoritative count still requires INSPIRE.

VERIFY

复核一组光学模型计算Cross-checking an optical-model calculation

任务是计算 50 MeV 的 n+⁹⁰Zr 弹性散射，使用 KD02 全局光学势。参数直接取自本地保存的 Koning 原始 kd02.f，没有凭记忆重写公式。两套独立实现一致到 8 位有效数字。 n+⁹⁰Zr elastic scattering at 50 MeV with the KD02 global optical potential. The parameters were not rewritten from memory: the run reused Koning's own kd02.f already on disk, and the two independent implementations agreed to 8 significant figures.

这里有个很隐蔽的坑。FRESCO 使用 R = r₀(Aₚ1/3 + At1/3)，KD02 则把半径定义为 R = r₀At1/3。如果没有设置 ap=0，所有半径都会大 22%，程序照样运行，截面看上去也很合理。最终得到 σR = 1301.640 mb。积分步长连续减半两次后稳定到 9 位有效数字，匹配半径和分波数加倍也不改变结果。再用独立求解器互检，COLOSS（复标度 Lagrange-Laguerre）给出 1299.188 mb，FRESCO（Numerov 耦合道）给出 1299.191 mb，两者一致到 6 位有效数字。设 W = 0 后吸收精确归零，通量守恒。1301.640 和 1299.19 并不矛盾，差别只在质量约定：前者用物理质量（1.008665 和 89.904698），互检那一对用整数质量（A = 1 和 A = 90），同一套 FRESCO 输入卡换掉这两个数就能在两者之间来回复现，与求解器无关。 It caught a trap that fails silently: FRESCO builds radii as R = r₀(Aₚ1/3 + At1/3) while KD02 is defined on R = r₀At1/3. Without ap=0 every radius is 22% too large and the resulting cross section still looks entirely plausible. Final σR = 1301.640 mb: stable to 9 significant figures under two halvings of the radial step and unchanged when the matching radius and partial-wave limit were doubled. A second, structurally unrelated solver then checked it: COLOSS (complex-scaled Lagrange-Laguerre) gave 1299.188 mb against FRESCO's 1299.191 mb, 6 significant figures. With W = 0 the absorption vanished exactly and flux conservation held. The two figures do not disagree: 1301.640 uses physical masses (1.008665 and 89.904698) while the cross-check pair uses integer ones (A = 1 and A = 90). Swapping those two numbers in one deck moves the answer between them, so the gap belongs to the mass convention and not to either solver.

实验数据：EXFOR 没有 50 MeV 的 n+⁹⁰Zr 弹性散射测量。于是换到两个有数据的能量重新计算，两次都没有自由参数，两个能量报同一个统计量：calc/data 的均值在 24 MeV 是 0.929，在 55 MeV 是 0.995（中位比分别为 0.891 和 0.961）。下图是计算与数据的直接比较。 Experimental data: EXFOR contains no n+⁹⁰Zr elastic-scattering measurement at 50 MeV. The calculation was repeated at the two measured energies, with no free parameters and the same statistic reported for both panels: the mean calc/data ratio is 0.929 at 24 MeV and 0.995 at 55 MeV (medians 0.891 and 0.961). The figure compares those calculations with the data.

KD02 global optical potential versus EXFOR measurements for neutron elastic scattering on zirconium, at 24 MeV and 55 MeV
绿线是 KD02 全局光学势的零自由参数预测，粉点是 EXFOR 数据。两个面板都按各自的实验能量重新计算。左图为 24 MeV 的富集 ⁹⁰Zr 全角度数据（Wang & Rapaport，EXFOR 13160.004），右图为 55 MeV 的天然 Zr 前角数据（Ibaraki TIARA，EXFOR 22480.005）。50 MeV 没有测量，所以不在图中。 Green shows the zero-free-parameter prediction of the KD02 global optical potential; pink points are EXFOR measurements. Each panel was computed at its measured energy. Left: 24 MeV, enriched ⁹⁰Zr, full angular range (Wang & Rapaport, EXFOR 13160.004). Right: 55 MeV, natural Zr, forward angles only (Ibaraki TIARA, EXFOR 22480.005). There is no 50 MeV panel because no measurement exists at that energy.

语料地图 · Corpus mapThe corpus map

55,850 篇论文的引用投影 A citation projection of 55,850 papers

这张交互地图是 61,171 篇论文语料的静态快照，其中 55,850 篇达到引用度数门槛。当前公开仓库附带 61,059 个论文页，两批数据的生成时间和筛选范围不同。引用关系决定论文的位置，地形高低表示论文密度，地名只标在明显聚集的 PhySH 主题上。引用边参与布局计算，但不画在图上，否则大量连线会盖住地图。 This interactive map is a static snapshot of a 61,171-paper corpus; 55,850 papers pass the citation-degree threshold. The public repository currently ships 61,059 paper pages, produced at a different time and with different selection rules. Citation structure sets each position, terrain shows paper density, and place names mark PhySH topics that cluster in the projection. Citation edges constrain the layout but are not drawn, because the lines would obscure the map at this scale.

FUSION corpus citation map →打开交互地图 · 悬停查看论文Enter the interactive map · hover any paper

上面那块玻璃是折射出来的，不是磨砂：它把下面的论文密度压弯，而不是糊掉。指针移到图上，它会跟着走。The pane above is refracted, not frosted: it bends the paper density underneath instead of blurring it away. Move the pointer onto the map and it follows.

坐标来自引用关系Coordinates come from citations 坐标按引用邻接 → SVD → t-SNE 的流程计算。同主题论文的空间离散度中位数为全局的 0.66，最集中的主题低至 0.34。The layout uses citation adjacency → SVD → t-SNE. Median same-topic dispersion is 0.66 of the global value; the tightest topic reaches 0.34.
只标注聚集的主题Labels require spatial clustering 没有明显空间聚集的通用主题不标地名。地图黑白打印仍然清楚，常见色觉缺陷的读者也能分辨。Broad topics without a spatial cluster are not labeled. The map remains legible in grayscale and under common forms of color-vision deficiency.
整张地图只有一个文件The map is one file 所有数据都内嵌在页面中，不用 CDN，除访问统计外不发起外部请求。打开文件即可浏览，无须构建。All data are inlined. The page uses no CDN and no build step; its only external request is the visit counter.

三维核素图 · Nuclide chartThe nuclide chart

3558 个基态，一片同量异位素山谷 3558 ground states as an isobaric valley

NUBASE2020 的全部基态画成三维柱图。柱高是 ΔE = M(Z,A) − minZ M(Z,A)，稳定核落在谷底，每个 A 的截面就是一条质量抛物线。颜色可以切换为主衰变道、半衰期或 B/A；时间轴从 1896 年走到 2020 年，逐年点亮每个同位素被鉴别出来的时刻。 Every NUBASE2020 ground state as a 3D column. Height is ΔE = M(Z,A) − minZ M(Z,A), so stable nuclei sit on the valley floor and each A-slice is a mass parabola. Colour switches between decay mode, half-life and B/A, and a timeline from 1896 to 2020 lights up each isotope in the year it was identified.

3D nuclide chart around lead, N = 126 →打开三维核素图 · 拖动旋转Enter the nuclide chart · drag to orbit

开始使用 · Get startedGet started

三步，大约五分钟 Three steps, about five minutes

不用提前安装核物理程序。第一次调用时，FUSION 会自行下载并编译。下面的命令已经在 macOS 和 Linux 上跑过。 No nuclear-physics code needs to be installed in advance. FUSION downloads and builds a code when it is first used. The commands below have been run on macOS and Linux.

1下载 FUSION 仓库Clone FUSION

仓库约 229 MB，主要内容是 61,167 页离线文献库。The repository is about 229 MB, mostly the 61,167-page offline literature base.

git clone https://github.com/jinleiphys/FUSION.git
cd FUSION

2下载 fusion 命令行程序Download the fusion CLI

把命令行程序放进刚才的目录。下面下载的是 Apple 芯片 Mac 版本；Intel Mac 请改用 darwin-x64，Linux 请改用 linux-x64。其他平台可在 releases 页找到。Download the command-line program into the same directory. The command below uses the Apple-silicon package; use darwin-x64 for Intel Macs and linux-x64 for Linux. Other builds are listed on the releases page.

curl -fsSL https://github.com/jinleiphys/FUSION/releases/latest/download/fusion-darwin-arm64.tar.gz | tar -xz
# macOS 才需要这一行only
xattr -d com.apple.quarantine fusion

当前二进制文件没有 Apple 开发者签名，macOS 默认会拦截。第二条命令用来清除隔离标记。The current binary has no Apple developer signature, so macOS blocks it by default. The second command clears the quarantine flag.

3运行 FUSIONOpen it

./fusion

发出第一条消息后，FUSION 会询问是否配置模型、研究方向、配色和私人研究档案。研究档案会从你的论文中提取课题词、合作者和语料内引用关系。这一步不会打断当前任务，也只会出现一次；输入“跳过”即可关闭。After the first message, FUSION offers to configure a model, research areas, colours, and a private research profile. The profile extracts topics, co-authors, and in-corpus citations from your papers. Setup does not block the current task and appears only once; enter “skip” to decline.

常见问题Common questions

需要 API key 吗？去哪里申请？Do I need an API key, and where from? 需要。FUSION 本身不包含大模型，请向 DeepSeek、Qwen、GLM、Claude 或 GPT 的服务商申请 key。首次配置会运行 fusion auth login，按提示粘贴即可。Yes. FUSION does not include a model. Obtain a key from DeepSeek, Qwen, GLM, Claude, or GPT, then paste it when setup runs fusion auth login.
提示 command not found: fusioncommand not found: fusion 请使用 ./fusion。开头的点和斜杠表示运行当前目录中的程序。如果希望在任意目录直接输入 fusion，可执行 mkdir -p ~/.local/bin && mv fusion ~/.local/bin/。仍然找不到时，再执行 echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc，然后重新打开终端。Use ./fusion; the leading dot-slash selects the program in the current directory. To run fusion from any directory, use mkdir -p ~/.local/bin && mv fusion ~/.local/bin/. If the command is still unavailable, run echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc and open a new terminal.
安装 FRESCO 或 TALYS 会改动系统目录吗？Does installing FRESCO or TALYS modify system directories? 不会。程序都装在 ~/.cache/fusion/ 下，不写系统目录；删除这个目录即可移除。编译需要 gfortran 和 C++ 编译器，缺少时 FUSION 会给出提示。TALYS 约占 11 GB，其中 8.6 GB 是核结构数据库，安装前请留意磁盘空间。Programs are installed under ~/.cache/fusion/, not in system directories. Removing that directory uninstalls them. Compilation requires gfortran and a C++ compiler; FUSION reports missing tools. TALYS needs about 11 GB, including an 8.6 GB nuclear-structure database, so check disk space before installing it.
计算结果可以直接使用吗？Can I use a result without checking it? 不可以。基准测试只能证明程序安装正确，并且复现过已知结果，无法替你检查新任务的物理设定。每个技能都会报告复现精度和未验证的部分，物理判断仍由用户负责。No. A benchmark shows that the code was installed correctly and reproduced a known result; it does not validate the physics choices in a new task. Each skill reports its agreement and what remains untested. The user remains responsible for the physics.
出问题去哪里反馈？Something broke 请到仓库提交 issue。无论是程序装不上，还是结果看着合理却实际有错，都请附上输入、输出、操作系统和编译器版本。Open a repository issue. For a failed installation or a plausible but incorrect result, include the input, output, operating system, and compiler version.

测试版最需要哪些反馈Feedback needed for the beta

作者每天都在使用 v0.1.0，但还没有外部用户完成测试。目前最想知道下面这些问题。The author uses v0.1.0 daily, but no external user has completed a test. Please report these issues first.

  1. 看起来对、其实错的结果。A result that looked right and was wrong. 例如，半径约定写错的 FRESCO 输入卡仍能正常运行，截面却可能偏差 20%，而且不会报错。如果遇到这类结果，请提供输入卡、计算值和参考值。For example, a FRESCO deck with the wrong radius convention can run without an error while shifting the cross section by 20%. If FUSION produces such a result, send the deck, the calculated value, and the reference value.
  2. 装不上的程序。A code that will not install. 20 个程序中，目前只有 FRESCO 从空缓存开始完整测试过安装流程。请附上报错、操作系统和编译器版本。Of the 20 programs, only FRESCO has been installed in a test that began with an empty cache. Include the error, operating system, and compiler version.
  3. 任何用着别扭的地方。Anything that felt stupid. 首次配置目前只有作者测试过。步骤多余、提示难懂、默认值不合习惯，都请直接指出。The first-run flow has only been tested by its author. Report unnecessary steps, unclear prompts, and poor defaults.
  4. 你还希望支持哪个程序。Which code you wish had a skill.

TALYS 需要下载约 11 GB，不建议拿它做第一次测试。macOS 首次运行若被拦截，请执行 xattr -d com.apple.quarantine fusion。TALYS downloads about 11 GB, so it is not a good first test. If macOS blocks the first run, use xattr -d com.apple.quarantine fusion.

打开仓库Open the repository → 下载页Downloads →

项目进度 · RoadmapRoadmap

现在做到这里 Current project status

v0.1.0 已经公开，20 个程序技能和离线知识库都已随仓库发布。现在主要补全新机器上的安装测试，并根据外部用户的实际反馈修问题。 v0.1.0 is public, with 20 code skills and the offline knowledge base shipped in the repository. Current work focuses on cold-start installation tests and fixes driven by external-user feedback.

Phase 0 · 已通过passed

质量验收Quality gate

三项测试都通过：精确找到指定文献；独立生成 FRESCO 输入卡，结果一致到 4–5 位有效数字；10 条引文逐条核实无误。Checked against references: exact retrieval of a specified paper; an independently generated FRESCO deck agreeing to 4–5 significant figures; 10 citations verified.

Phase 1 · 核心已完成core complete

品牌分叉 + CIRebrand fork + CI

fusion-core 只保留品牌补丁，每周自动重基上游 opencode。桌面图标和完整发行流水线还在补。fusion-core keeps only the brand patch and rebases onto upstream opencode each week. Desktop icons and the complete release pipeline remain in progress.

Phase 2 · 核心已完成core complete

技能包Skill pack

仓库已有 26 个技能，其中 20 个负责具体程序。程序技能暂停扩张，下一步先看真实用户会用什么。The repository ships 26 skills, including 20 for specific codes. Code-skill expansion is paused until real usage shows what people need next.

Phase 3 · 核心已完成core complete

知识库Knowledge base

kb-wiki 现有 61,167 页，包括 61,059 个论文页和 108 个主题页。引用和语义关系已经写入，今后随新增论文更新。kb-wiki contains 61,167 pages: 61,059 paper pages and 108 topic pages. Citation and semantic links are included; later releases will update them as papers are added.

Phase 4 · 正在做in progress

公开测试Public beta

v0.1.0 已经发布。全新机器上的安装测试还不够，目前只有 FRESCO 真正从空缓存完整跑过安装流程。v0.1.0 is released. Cold-start installation coverage is still thin; only FRESCO has completed the full install path from an empty cache.

FUSION

开源 · Open sourceOpen source

现在就能试 Available for testing now

作者已经在日常研究中使用这个版本，外部测试才刚开始。克隆仓库并下载命令行程序后，26 个技能和离线知识库会自动加载。最需要反馈的是安装失败，以及那些看着合理却实际有错的结果。 The author uses this version in daily research; external testing has just begun. After cloning the repository and downloading the CLI, all 26 skills and the offline knowledge base load automatically. Installation failures and plausible but incorrect results are the highest-priority reports.

试用 / 提交问题Try it / report an issue →