diff --git a/README.md b/README.md
index 9033e03..2c4497b 100644
--- a/README.md
+++ b/README.md
@@ -33,7 +33,7 @@
[](https://www.instagram.com/mr.abstractor_ust/)
[](https://www.linkedin.com/in/hongjinhe-hkust-edu)
-**[English](README.md) | [中文文档](README_CN.md)**
+**[English](README.md) | [中文文档](README_CN.md) | [日本語](docs/i18n/README_JA.md) | [한국어](docs/i18n/README_KO.md)**
**Alpha Flow Research · HongJin HE · HKUST / Stanford IHP · July 2026**
@@ -89,31 +89,37 @@ python demo/global_demo.py # reproduce this GIF end-to-end (~2 min, CPU, no
*— Alpha Flow Research · a Stanford basement · July 2026*
+*The longer version — how a world-model skeptic was converted by robots that dream, cars that predict, and a factor zoo that violates first principles — is **[docs/JOURNEY.md](docs/JOURNEY.md)**.*
+
---
## Table of Contents
1. [The Problem Nobody Has Solved](#the-problem-nobody-has-solved)
-2. [Why This Question Is Urgent — Right Now](#why-this-question-is-urgent--right-now)
-3. [What Came Before — And Why It Falls Short](#what-came-before--and-why-it-falls-short)
-4. [Two Kinds of World Model: Type 1 and Type 2](#two-kinds-of-world-model-type-1-and-type-2)
-5. [The Market as a Four-Level Game](#the-market-as-a-four-level-game)
-6. [The E-Game-C Architecture](#the-e-game-c-architecture)
-7. [The Mathematical Framework](#the-mathematical-framework-two-threads-one-theory)
-8. [The Unified Evolution Equation](#the-unified-evolution-equation-theorem-91)
-9. [The Seven Theorems](#the-seven-theorems)
-10. [Reflexivity — Soros, Formalized](#reflexivity--soros-formalized)
-11. [The Agent Taxonomy](#the-agent-taxonomy)
-12. [Connection to the 2026 Fields Medal](#connection-to-the-2026-fields-medal-deng-yu-邓煜)
-13. [What This Makes Possible](#what-this-makes-possible)
-14. [Data Requirements & Research Roadmap](#data-requirements--research-roadmap)
-15. [The 17-Day Notebook Series](#the-17-day-notebook-series)
-16. [Repository Structure](#repository-structure)
-17. [Quick Start](#quick-start)
-18. [Related Work & Positioning](#related-work--positioning)
-19. [Project Roadmap](#project-roadmap)
-20. [References](#references)
-21. [Citation](#citation)
+2. [The First-Principles Bet — Causality over Correlation](#the-first-principles-bet--causality-over-correlation)
+3. [Why This Question Is Urgent — Right Now](#why-this-question-is-urgent--right-now)
+4. [What Came Before — And Why It Falls Short](#what-came-before--and-why-it-falls-short)
+5. [Two Kinds of World Model: Type 1 and Type 2](#two-kinds-of-world-model-type-1-and-type-2)
+6. [The Market as a Four-Level Game](#the-market-as-a-four-level-game)
+7. [The E-Game-C Architecture](#the-e-game-c-architecture)
+8. [The Mathematical Framework](#the-mathematical-framework-two-threads-one-theory)
+9. [The Unified Evolution Equation](#the-unified-evolution-equation-theorem-91)
+10. [The Seven Theorems](#the-seven-theorems)
+11. [Reflexivity — Soros, Formalized](#reflexivity--soros-formalized)
+12. [The Agent Taxonomy](#the-agent-taxonomy)
+13. [Connection to the 2026 Fields Medal](#connection-to-the-2026-fields-medal-deng-yu-邓煜)
+14. [What This Makes Possible](#what-this-makes-possible)
+15. [Two Products, One Engine](#two-products-one-engine)
+16. [Data Requirements & Research Roadmap](#data-requirements--research-roadmap)
+17. [The 17-Day Notebook Series](#the-17-day-notebook-series)
+18. [Repository Structure](#repository-structure)
+19. [Quick Start](#quick-start)
+20. [Related Work & Positioning](#related-work--positioning)
+21. [Project Roadmap — Phase 1 and Phase 2](#project-roadmap--phase-1-and-phase-2)
+22. [Star History](#star-history)
+23. [Partnerships & Contact](#partnerships--contact)
+24. [References](#references)
+25. [Citation](#citation)
---
@@ -133,6 +139,28 @@ The answer is yes. This repository presents that theory, and its engineering imp
---
+## The First-Principles Bet — Causality over Correlation
+
+Every mainstream quant methodology — factor mining, time-series modelling, feature engineering — rests on one unexamined premise: **that historical data will repeat.** It will not. Every backtested pattern is a snapshot of a game whose players have since changed their strategies, partly *because* the pattern was found. That is the Lucas critique with teeth, and it is why the factor zoo decays and deployed models die.
+
+But something *does* repeat. Not the data — the **world that generates the data**: institutions with mandates, regulators with rules, incentives that do not change when a signal is published. History does not repeat its prices; it repeats its *mechanisms*. So the only factors that can carry meaning are the ones derived by going back to the world itself — and this project refuses to model the residue when it can model the generator:
+
+| | Pattern paradigm (factors, TS-ML) | World-model paradigm (this repo) |
+|---|---|---|
+| **Object modeled** | The data the market left behind | The market itself: agents, constraints, equilibrium |
+| **A "factor" is** | A correlation that worked in-sample | A term in an equilibrium condition, with a mechanism attached |
+| **When regimes break** | Silent failure — nothing to explain with | The explanation *is* the model: which agents, which constraint, which coupling |
+| **Deployment** | Erodes the signal (crowding) | Reinforces the signal (equilibria are self-consistent) |
+
+Two consequences follow, and they are this project's identity:
+
+1. **Explainability is the direction of inference, not a feature.** We do not fit prices and hope for meaning; we model the mechanism and *derive* what prices must do. Every output is attributable to named agents, named constraints, and an equilibrium condition you can inspect.
+2. **The model survives contact with the market.** A prediction that is an equilibrium does not evaporate when acted on — being acted on is how equilibria assert themselves.
+
+> **We did not find structure in the data. We modeled the structure that makes the data.** That sentence is the entire difference between this repository and every pattern-mining stack in production today.
+
+---
+
## Why This Question Is Urgent — Right Now
Three simultaneous forces are breaking the old paradigm faster than at any point in history:
@@ -753,6 +781,38 @@ When retail investors coordinate (GME, AMC, any future short squeeze), the behav
---
+## Two Products, One Engine
+
+The same E-Game-C core surfaces as two products — one per audience.
+
+### To-C · The denoised equilibrium price (mid/long-horizon)
+
+Theorem 1 says the behavioral noise ν_η cannot be out-traded — by anyone, retail least of all. What a long-horizon investor actually needs is the component *underneath* the noise: the **equilibrium track P^eq** — what the asset is worth once every agent has played its rational strategy — and the divergence **D_t = P_t/P^eq_t − 1**, which answers exactly one question: *are you buying value, or buying crowding?*
+
+[](demo/denoised_price_2026.py)
+
+The scenario above is **synthetic** — a concept demo shaped on the July 2026 memory-sector unwind, the month the Philadelphia Semiconductor Index lost 19% (its worst since 2008) while retail flows capitulated at lows that institutions were buying. In the simulation, the signal — divergence above threshold *while the institutional mean field rotates out* — fires **11 trading days (≈2 weeks) before the unwind accelerates**. The real-data version, under the same honesty rules as the 2008 hindcast (walk-forward, no look-ahead), is specified as experiment **[E7](RESOURCES.md)**.
+
+```bash
+python demo/denoised_price_2026.py # reproduce the figure — synthetic, CPU, no keys
+```
+
+> *Research signal, not advice.* The denoised price is an instrument-level research layer for mid/long-horizon positioning. It is not personalized investment advice, and this repository is not an advisor.
+
+### To-B · Structural risk for institutions
+
+For funds, risk desks, and platforms, the same machinery runs in the other direction — selling not the equilibrium but the **distance from it**:
+
+- **Λₜ regime monitoring** — the 2008/2020-grade basin-exit signal, intraday ([`online/regime_detector.py`](online/regime_detector.py))
+- **Crowding decomposition** — how much of a book's P&L is equilibrium drift vs behavioral wedge, per position
+- **Event-operator scenario analysis** — M&A, policy moves, delistings applied as first-class operators to *today's* state, not as historical analogies
+
+The production loop (Airflow DAG → noise → encoder → MFG → signal → execution) is already scaffolded in [`online/`](online/) with an Alpaca paper-trading stub. This is Horizon 3 of the roadmap.
+
+**One engine, two surfaces:** the retail product is the equilibrium; the institutional product is the deviation from it. Both are outputs of the same solve.
+
+---
+
## Data Requirements & Research Roadmap
The mathematics is closed and the code runs end-to-end on synthetic data. The road from prototype to validated instrument is **data** — hundreds of fragmented streams, each feeding a specific term of a specific equation. The complete acquisition plan lives in **[DATA_REQUIREMENTS.md](DATA_REQUIREMENTS.md)**: every source named, every stub located, every cost tiered, and — for the data that exists nowhere — the experiments that create it.
@@ -847,6 +907,8 @@ MicroWorld/
├── demo/
│ ├── run_egamec.py # 30-second E-Game-C pipeline demo
│ ├── global_demo.py # ★ The animated world-model demo (GIF above)
+│ ├── hindcast_2008.py # The 2008 walk-forward hindcast (figure at top)
+│ ├── denoised_price_2026.py # To-C concept demo: denoised equilibrium price
│ └── synthetic_market.py # Dual-noise synthetic market generator
│
├── scripts/make_figures.py # Regenerates every README figure from library code
@@ -854,6 +916,8 @@ MicroWorld/
├── figures/ # All SVG diagrams + generated PNGs + demo GIF
├── tests/ # 50 tests, all passing (noise · events · features)
├── DATA_REQUIREMENTS.md # ★ The complete data & experiment roadmap
+├── RESOURCES.md # ★ Data + compute, priced — the use-of-funds view
+├── docs/ # JOURNEY · PHASE2_NEURAL_GAME · ONE_PAGER · i18n
├── CITATION.cff # Citable metadata
└── .github/workflows/ci.yml # CI: pytest on 3.11 / 3.12
```
@@ -866,6 +930,10 @@ MicroWorld/
|---|---|---|
| **Engineering Implementation** (E-Game-C) | [us-equity-world-model](https://github.com/hongjin-he/us-equity-world-model) | Full build manual: data layer, encoder, MFG solver, controller, backtest, deployment |
| **Mathematical Paper** | [mathmatical-framework-for-world-models-in-quant-finance](https://github.com/hongjin-he/mathmatical-framework-for-world-models-in-quant-finance) | Alpha Flow 02: all proofs, 9 theorems, 25 pages |
+| **The Journey** | [docs/JOURNEY.md](docs/JOURNEY.md) | From world-model skeptic to this architecture — with the robots, cars, and dreams that did the converting |
+| **Phase 2 Design** | [docs/PHASE2_NEURAL_GAME.md](docs/PHASE2_NEURAL_GAME.md) | The Neural Network Game Structure: every neuron an agent-network |
+| **Resources & Use of Funds** | [RESOURCES.md](RESOURCES.md) | Data + compute, priced by scenario; experiments E7–E11 |
+| **One-Pager** | [docs/ONE_PAGER.md](docs/ONE_PAGER.md) | The whole project on one page, for partners and investors |
---
@@ -882,6 +950,9 @@ python demo/global_demo.py
# 2 · The 30-second pipeline demo
python demo/run_egamec.py
+# 2b · The denoised-price concept demo (To-C product line, synthetic)
+python demo/denoised_price_2026.py
+
# 3 · Regenerate every figure in this README from library code
python scripts/make_figures.py
@@ -934,7 +1005,7 @@ Where MicroWorld sits relative to each adjacent literature:
---
-## Project Roadmap
+## Project Roadmap — Phase 1 and Phase 2
The project holds itself to a **dual standard**, on purpose:
@@ -951,6 +1022,12 @@ E3 (equilibrium consistency vs 13F/COT), E6 (L0 transmission), encoder trained o
**Horizon 3 — the industrial product (P2).**
Live daily pipeline (Airflow DAG already scaffolded), paper-trading via Alpaca stub, dashboard, and — compute permitting — the Type 2 sandbox with the Type 1 equilibrium as its outer loop.
+---
+
+Everything above is **Phase 1 — the game-theoretic core**: mathematical structure standing in for scarce data, and not as a compromise — Theorem 1 proves part of the residual can *only* be explained by mechanism, never estimated away, so structure-first is the theoretically correct regime at today's data access. When the cleaned panel and H200-class compute both exist, the architecture is scheduled to shed its skin:
+
+**Phase 2 — the [Neural Network Game Structure](docs/PHASE2_NEURAL_GAME.md).** A network in which **every neuron is itself a small neural network** — one per institution or retail cohort, with its own objective and information set; regulatory reality imposed as architecture (all agents of a type share their regulator's constraint module); the forward pass *is* the game being played. Whether the environment and the population live in one network or two is assigned to experiment, not taste (E8). Training difficulty roughly squares — which is exactly why the design document is public now and the training waits for the hardware. Triggers, pilots, and budget: [RESOURCES.md](RESOURCES.md).
+
Contributions welcome — see [CONTRIBUTING.md](CONTRIBUTING.md).
---
@@ -1013,6 +1090,43 @@ Contributions welcome — see [CONTRIBUTING.md](CONTRIBUTING.md).
[39] J. Sirignano, K. Spiliopoulos, "DGM: a deep learning algorithm for solving partial differential equations," *J. Computational Physics*, 2018.
[40] R. Cont, J.-P. Bouchaud, "Herd behavior and aggregate fluctuations in financial markets," *Macroeconomic Dynamics*, 2000.
+**World models across domains, and the factor zoo (the journey — [docs/JOURNEY.md](docs/JOURNEY.md))**
+[41] P. Wu, A. Escontrela, D. Hafner, P. Abbeel, K. Goldberg, "DayDreamer: world models for physical robot learning," *CoRL*, 2022.
+[42] A. Hu et al., "GAIA-1: a generative world model for autonomous driving," arXiv:2309.17080, 2023.
+[43] J. Bruce et al., "Genie: generative interactive environments," *ICML*, 2024.
+[44] NVIDIA, "Cosmos world foundation model platform for physical AI," arXiv:2501.03575, 2025.
+[45] M. Assran et al., "V-JEPA 2: self-supervised video models enable understanding, prediction and planning," arXiv:2506.09985, 2025.
+[46] J. H. Cochrane, "Presidential address: discount rates," *Journal of Finance*, 2011.
+
+---
+
+## Star History
+
+Stars are this project's market validation — public, timestamped, and unfakeable in slope. The curve updates live:
+
+[](https://star-history.com/#hongjin-he/MicroWorld&Date)
+
+If the framework earned your star, the next-highest-leverage contribution is one issue: tell us which claim you'd attack first.
+
+---
+
+## Partnerships & Contact
+
+MicroWorld is looking for exactly four kinds of counterparty:
+
+| Who | What we bring | What we need |
+|---|---|---|
+| **Academic research groups** | A NeurIPS-workshop-ready experiment suite (E1–E7) executable in one quarter | WRDS-grade data access; co-authorship welcome |
+| **Quant funds & prop desks** | The Λₜ early-warning and crowding-decomposition line (To-B) — plus first access to E7 results | A pilot conversation and honest adversarial review |
+| **Compute partners** | A publicly specified Phase 2 architecture ([design doc](docs/PHASE2_NEURAL_GAME.md)) that is H200-shaped by construction | H100/H200 hours for experiments E8–E11 |
+| **Retail platforms & media** | The denoised-price research layer (To-C) — a story retail investors actually need after July 2026 | Distribution and product feedback |
+
+Start with the **[one-pager](docs/ONE_PAGER.md)**; budgets and scenarios are in **[RESOURCES.md](RESOURCES.md)**.
+
+**Channels:** [LinkedIn](https://www.linkedin.com/in/hongjinhe-hkust-edu) · [X](https://x.com/Mr_Abstractor) · [GitHub issues](https://github.com/hongjin-he/MicroWorld/issues) — technical objections get the fastest replies.
+
+> *Compliance note: this repository is research software. Nothing in it is investment advice, portfolio management, or an offer of securities.*
+
---
## Citation
diff --git a/README_CN.md b/README_CN.md
index ce417c5..e583c12 100644
--- a/README_CN.md
+++ b/README_CN.md
@@ -27,7 +27,7 @@
[](https://www.instagram.com/mr.abstractor_ust/)
[](https://www.linkedin.com/in/hongjinhe-hkust-edu)
-**[English](README.md) | [中文文档](README_CN.md)**
+**[English](README.md) | [中文文档](README_CN.md) | [日本語](docs/i18n/README_JA.md) | [한국어](docs/i18n/README_KO.md)**
**Alpha Flow Research · 何泓锦 · 香港科技大学 / 斯坦福 IHP · 2026年7月**
@@ -83,31 +83,37 @@ python demo/global_demo.py # 端到端复现此 GIF（约 2 分钟，仅 CPU
*—— Alpha Flow Research · 斯坦福某间地下室 · 2026年7月*
+*更长的版本——一个世界模型怀疑论者，如何被会做梦的机器人、会预判的自动驾驶和一个违背第一性原理的因子动物园说服——在 **[docs/JOURNEY.md](docs/JOURNEY.md)**（英文）。*
+
---
## 目录
1. [没有人真正解决的问题](#没有人真正解决的问题)
-2. [为什么这个问题现在最紧迫](#为什么这个问题现在最紧迫)
-3. [前人的研究——以及它们为什么不够](#前人的研究以及它们为什么不够)
-4. [两种世界模型：Type 1 与 Type 2](#两种世界模型type-1-与-type-2)
-5. [市场作为四层博弈](#市场作为四层博弈)
-6. [E-Game-C 架构](#e-game-c-架构)
-7. [数学框架：双线并行](#数学框架双线并行)
-8. [统一演化方程](#统一演化方程定理91)
-9. [七大定理](#七大定理)
-10. [反身性——形式化索罗斯](#反身性形式化索罗斯)
-11. [主体分类法](#主体分类法)
-12. [与2026年菲尔兹奖的联系](#与2026年菲尔兹奖的联系邓煜)
-13. [这让什么成为可能](#这让什么成为可能)
-14. [数据需求与研究路线图](#数据需求与研究路线图)
-15. [17 天笔记本系列](#17-天笔记本系列)
-16. [仓库结构](#仓库结构)
-17. [快速开始](#快速开始)
-18. [相关工作与定位](#相关工作与定位)
-19. [项目路线图](#项目路线图)
-20. [参考文献](#参考文献)
-21. [引用](#引用)
+2. [第一性原理之赌——因果而非相关](#第一性原理之赌因果而非相关)
+3. [为什么这个问题现在最紧迫](#为什么这个问题现在最紧迫)
+4. [前人的研究——以及它们为什么不够](#前人的研究以及它们为什么不够)
+5. [两种世界模型：Type 1 与 Type 2](#两种世界模型type-1-与-type-2)
+6. [市场作为四层博弈](#市场作为四层博弈)
+7. [E-Game-C 架构](#e-game-c-架构)
+8. [数学框架：双线并行](#数学框架双线并行)
+9. [统一演化方程](#统一演化方程定理91)
+10. [七大定理](#七大定理)
+11. [反身性——形式化索罗斯](#反身性形式化索罗斯)
+12. [主体分类法](#主体分类法)
+13. [与2026年菲尔兹奖的联系](#与2026年菲尔兹奖的联系邓煜)
+14. [这让什么成为可能](#这让什么成为可能)
+15. [同一引擎，两条产品线](#同一引擎两条产品线)
+16. [数据需求与研究路线图](#数据需求与研究路线图)
+17. [17 天笔记本系列](#17-天笔记本系列)
+18. [仓库结构](#仓库结构)
+19. [快速开始](#快速开始)
+20. [相关工作与定位](#相关工作与定位)
+21. [项目路线图——第一阶段与第二阶段](#项目路线图第一阶段与第二阶段)
+22. [Star 增长曲线](#star-增长曲线)
+23. [合作与联系](#合作与联系)
+24. [参考文献](#参考文献)
+25. [引用](#引用)
---
@@ -127,6 +133,28 @@ python demo/global_demo.py # 端到端复现此 GIF（约 2 分钟，仅 CPU
---
+## 第一性原理之赌——因果而非相关
+
+每一种主流量化方法——因子挖掘、时间序列建模、特征工程——都建立在一个未经审视的前提上：**历史数据会重演。** 它不会。每一个回测出来的模式，都是一场博弈的快照，而博弈的玩家早已改变了策略——部分原因恰恰是这个模式被发现了。这就是长出牙齿的卢卡斯批判，也是因子动物园持续衰减、已部署模型不断死亡的原因。
+
+但确实有东西在重演。不是数据——而是**生成数据的那个世界**：有法定授权的机构、有规则的监管者、不因信号被发表而改变的激励结构。历史不会重复它的价格，历史重复的是它的*机制*。因此，唯一能承载意义的因子，是回到世界本身推导出来的因子——本项目拒绝为残渣建模，因为我们可以为生成器建模：
+
+| | 模式范式（因子、时序ML） | 世界模型范式（本仓库） |
+|---|---|---|
+| **建模对象** | 市场留下的数据 | 市场本身：主体、约束、均衡 |
+| **一个"因子"是** | 一条样本内有效的相关性 | 均衡条件中的一项，附带机制 |
+| **机制断裂时** | 静默失效——无从解释 | 解释*就是*模型：哪些主体、哪条约束、哪个耦合 |
+| **部署之后** | 信号被侵蚀（拥挤） | 信号被强化（均衡自洽） |
+
+由此推出两个结论，它们就是本项目的身份：
+
+1. **可解释性不是附加功能，而是推断的方向。** 我们不是拟合价格再祈祷其中有意义；我们建模机制，然后*推导*价格必须如何运动。每一个输出都可归因于具名的主体、具名的约束和一个可检查的均衡条件。
+2. **模型经得起与市场的接触。** 一个本身就是均衡的预测，不会因为被执行而蒸发——被执行，正是均衡实现自身的方式。
+
+> **我们没有在数据里找到结构。我们建模了生成数据的结构。** 这一句话，就是本仓库与今天所有在产模式挖掘系统的全部区别。
+
+---
+
## 为什么这个问题现在最紧迫
三股力量正在以历史上最快的速度瓦解旧范式：
@@ -734,6 +762,38 @@ $\pi^* = (\gamma\Sigma + \lambda I)^{-1}(\mu + \lambda\,\mu^{\text{MFG}})$
---
+## 同一引擎，两条产品线
+
+同一个 E-Game-C 内核，面向两类受众呈现为两个产品。
+
+### To-C · 去噪均衡价格（中长期持有）
+
+定理1说明：行为噪声 ν_η 是任何人都无法通过交易战胜的——散户尤其如此。一个长期投资者真正需要的，是噪声*之下*的那个分量：**均衡轨迹 P^eq**——当每一个主体都打出理性策略后，这个资产值多少钱——以及偏离度 **D_t = P_t/P^eq_t − 1**。它只回答一个问题：*你买入的是价值，还是拥挤？*
+
+[](demo/denoised_price_2026.py)
+
+上图是**合成数据**——一个以2026年7月存储板块崩盘为情境原型的概念演示：那个月费城半导体指数下跌19%（2008年以来最差单月），散户在低点恐慌抛售，而机构在同一低点买入。在模拟中，信号——偏离度突破阈值*且机构平均场正在撤出*——在崩盘加速前 **11个交易日（约两周）** 触发。真实数据版本遵循与2008年回测完全相同的诚实规则（滚动前向、无未来函数），已定义为实验 **[E7](RESOURCES.md)**。
+
+```bash
+python demo/denoised_price_2026.py # 复现上图——合成数据，CPU，无需密钥
+```
+
+> *研究信号，而非投资建议。* 去噪价格是面向中长期持仓的标的级研究层。它不是个性化投资建议，本仓库也不是投资顾问。
+
+### To-B · 机构级结构性风险
+
+对基金、风控台和平台，同一套均衡机器反向运行——卖的不是均衡本身，而是**与均衡的距离**：
+
+- **Λₜ 机制监控**——2008/2020级别的稳定域逃逸信号，盘中运行（[`online/regime_detector.py`](online/regime_detector.py)）
+- **拥挤度分解**——一本账簿的盈亏中，多少来自均衡漂移、多少来自行为楔子，逐仓位归因
+- **事件算子情景分析**——并购、政策、退市作为第一等算子作用于*今天*的状态，而非历史类比
+
+生产环闭环（Airflow DAG → 噪声 → 编码器 → MFG → 信号 → 执行）已在 [`online/`](online/) 搭好脚手架，含 Alpaca 模拟盘 stub。这是路线图的地平线3。
+
+**同一引擎，两个表面：** 散户产品卖的是均衡，机构产品卖的是对均衡的偏离。二者是同一次求解的两个输出。
+
+---
+
## 数据需求与研究路线图
数学已闭合，代码在合成数据上端到端运行。从原型到经过验证的研究工具，路上缺的是**数据**——数百条零散的数据流，每一条喂给特定方程的特定一项。完整获取计划在 **[DATA_REQUIREMENTS.md](DATA_REQUIREMENTS.md)**：每个数据源具名、每个 stub 定位到行、每项成本分档，以及——对于世界上还不存在的数据——创造它们的实验设计。
@@ -828,6 +888,8 @@ MicroWorld/
├── demo/
│ ├── run_egamec.py # 30秒 E-Game-C 流水线演示
│ ├── global_demo.py # ★ 动画世界模型演示（顶部GIF）
+│ ├── hindcast_2008.py # 2008滚动前向回测（顶部图）
+│ ├── denoised_price_2026.py # To-C概念演示：去噪均衡价格
│ └── synthetic_market.py # 双噪声合成市场生成器
│
├── scripts/make_figures.py # 从库代码重新生成README全部图表
@@ -835,6 +897,8 @@ MicroWorld/
├── figures/ # 全部SVG图 + 生成的PNG + 演示GIF
├── tests/ # 50个测试，全部通过（噪声·事件·特征）
├── DATA_REQUIREMENTS.md # ★ 完整数据与实验路线图
+├── RESOURCES.md # ★ 数据+算力定价——资金用途视角
+├── docs/ # JOURNEY · PHASE2_NEURAL_GAME · ONE_PAGER · i18n
├── CITATION.cff # 可引用元数据
└── .github/workflows/ci.yml # CI：pytest on 3.11 / 3.12
```
@@ -847,6 +911,10 @@ MicroWorld/
|---|---|---|
| **工程实现**（E-Game-C） | [us-equity-world-model](https://github.com/hongjin-he/us-equity-world-model) | 完整构建手册：数据层、编码器、MFG求解器、控制器、回测、部署 |
| **数学论文** | [mathmatical-framework-for-world-models-in-quant-finance](https://github.com/hongjin-he/mathmatical-framework-for-world-models-in-quant-finance) | Alpha Flow 02：全部证明，9个定理，25页 |
+| **心路历程** | [docs/JOURNEY.md](docs/JOURNEY.md) | 从世界模型怀疑论者到这套架构——以及完成说服的那些机器人、汽车与梦境 |
+| **第二阶段设计** | [docs/PHASE2_NEURAL_GAME.md](docs/PHASE2_NEURAL_GAME.md) | 神经网络博弈结构：每个神经元本身是一个主体网络 |
+| **资源与资金用途** | [RESOURCES.md](RESOURCES.md) | 数据+算力按场景定价；实验E7–E11 |
+| **一页纸** | [docs/ONE_PAGER.md](docs/ONE_PAGER.md) | 面向合作方与投资人的单页项目全景 |
---
@@ -863,6 +931,9 @@ python demo/global_demo.py
# 2 · 30秒流水线演示
python demo/run_egamec.py
+# 2b · 去噪价格概念演示（To-C产品线，合成数据）
+python demo/denoised_price_2026.py
+
# 3 · 从库代码重新生成本README全部图表
python scripts/make_figures.py
@@ -915,7 +986,7 @@ MicroWorld 相对每条相邻研究线的位置：
---
-## 项目路线图
+## 项目路线图——第一阶段与第二阶段
本项目刻意坚持**双重标准**：
@@ -932,6 +1003,12 @@ E3（均衡一致性 vs 13F/COT）、E6（L0传导）、在真实面板上训练
**地平线3——工业产品（P2）。**
每日实盘流水线（Airflow DAG已搭好脚手架）、经Alpaca stub的模拟盘交易、监控面板，以及——算力允许时——以 Type 1 均衡为外层循环的 Type 2 沙盒。
+---
+
+以上全部属于**第一阶段——博弈论内核**：用数学结构补偿稀缺的数据，而且这不是妥协——定理1证明了残差中有一部分*只能*被机制解释、永远无法靠堆数据估计掉，所以在今天的数据获取水平上，结构优先在理论上就是正确的选择。当清理好的数据面板与 H200 级算力同时就位，这套架构将按计划蜕皮：
+
+**第二阶段——[神经网络博弈结构](docs/PHASE2_NEURAL_GAME.md)（NNGS）。** 一个**每个神经元本身就是一个小神经网络**的网络——每个机构或散户群体一个，各有自己的目标函数与信息集；监管现实作为架构直接施加（同一类型的所有主体共享其监管者的约束模块）；前向传播*就是*博弈的进行。环境与博弈人群放在一个网络还是两个网络，交给实验裁决而非审美（E8）。训练难度大约平方化——这正是设计文档现在就公开、而训练等待硬件的原因。触发条件、试点与预算：[RESOURCES.md](RESOURCES.md)。
+
欢迎贡献——见 [CONTRIBUTING.md](CONTRIBUTING.md)。
---
@@ -994,6 +1071,43 @@ E3（均衡一致性 vs 13F/COT）、E6（L0传导）、在真实面板上训练
[39] J. Sirignano, K. Spiliopoulos, "DGM: a deep learning algorithm for solving partial differential equations," *J. Computational Physics*, 2018.
[40] R. Cont, J.-P. Bouchaud, "Herd behavior and aggregate fluctuations in financial markets," *Macroeconomic Dynamics*, 2000.
+**跨领域世界模型与因子动物园（心路历程——[docs/JOURNEY.md](docs/JOURNEY.md)）**
+[41] P. Wu, A. Escontrela, D. Hafner, P. Abbeel, K. Goldberg, "DayDreamer: world models for physical robot learning," *CoRL*, 2022.
+[42] A. Hu et al., "GAIA-1: a generative world model for autonomous driving," arXiv:2309.17080, 2023.
+[43] J. Bruce et al., "Genie: generative interactive environments," *ICML*, 2024.
+[44] NVIDIA, "Cosmos world foundation model platform for physical AI," arXiv:2501.03575, 2025.
+[45] M. Assran et al., "V-JEPA 2: self-supervised video models enable understanding, prediction and planning," arXiv:2506.09985, 2025.
+[46] J. H. Cochrane, "Presidential address: discount rates," *Journal of Finance*, 2011.
+
+---
+
+## Star 增长曲线
+
+Star 是这个项目的市场验证——公开、带时间戳、斜率无法造假。曲线实时更新：
+
+[](https://star-history.com/#hongjin-he/MicroWorld&Date)
+
+如果这套框架配得上你的 star，那么杠杆率第二高的贡献是提一个 issue：告诉我们你最想攻击哪一条主张。
+
+---
+
+## 合作与联系
+
+MicroWorld 正在寻找恰好四类对手方：
+
+| 对象 | 我们带来什么 | 我们需要什么 |
+|---|---|---|
+| **学术研究组** | 一季度内可执行、达到 NeurIPS workshop 标准的实验套件（E1–E7） | WRDS 级数据权限；欢迎共同署名 |
+| **量化基金与自营交易台** | Λₜ 预警与拥挤度分解产品线（To-B），外加 E7 结果的优先获取 | 一次试点对话与诚实的对抗性审查 |
+| **算力伙伴** | 一套公开成文、天生为 H200 而设计的第二阶段架构（[设计文档](docs/PHASE2_NEURAL_GAME.md)） | 用于实验 E8–E11 的 H100/H200 机时 |
+| **零售平台与媒体** | 去噪价格研究层（To-C）——2026年7月之后散户真正需要的那个故事 | 分发渠道与产品反馈 |
+
+从**[一页纸](docs/ONE_PAGER.md)**开始；预算与场景在 **[RESOURCES.md](RESOURCES.md)**。
+
+**渠道：** [LinkedIn](https://www.linkedin.com/in/hongjinhe-hkust-edu) · [X](https://x.com/Mr_Abstractor) · [GitHub issues](https://github.com/hongjin-he/MicroWorld/issues)——技术性质疑回复最快。
+
+> *合规声明：本仓库是研究软件。其中任何内容均不构成投资建议、投资组合管理或证券要约。*
+
---
## 引用
diff --git a/RESOURCES.md b/RESOURCES.md
new file mode 100644
index 0000000..9ed3483
--- /dev/null
+++ b/RESOURCES.md
@@ -0,0 +1,92 @@
+# RESOURCES — What This Project Needs, Priced
+
+Two things stand between the current repository and a validated instrument:
+**cleaned data** and **training compute**. This file prices both. It is
+simultaneously the project's internal shopping list and — for a potential
+partner or investor — the use-of-funds statement. Experiment-level detail
+for the data lives in [DATA_REQUIREMENTS.md](DATA_REQUIREMENTS.md); the
+Phase 2 architecture that consumes the compute lives in
+[docs/PHASE2_NEURAL_GAME.md](docs/PHASE2_NEURAL_GAME.md).
+
+---
+
+## 1 · Pillar one: cleaned, *typed* data
+
+The framework does not want "more data" — it wants **each agent type's
+behavior observed at the level that type actually acts**. That is a
+different shopping list from a factor shop's, organized by which term of
+the model each stream feeds:
+
+| Agent type / model term | Data | Source | Cost tier |
+|---|---|---|---|
+| Physical vs behavioral noise (Thm 1) | Intraday trades & quotes, 500 equities × 10 yr | WRDS TAQ / Polygon | academic access / ~$2k·yr |
+| Event operators (all 22) | Corporate actions, M&A, IPO, delistings, earnings | CRSP + SDC + I/B/E/S | academic access |
+| Institutional mean field μ* (L1–L2) | Quarterly 13F holdings, futures COT positioning | EDGAR (free) + CFTC (free) | free |
+| Retail cohort policies (Day 16) | **The E5 LLM query atlas** — stratified audit of consumer AI investment advice; the dataset exists nowhere and we create it | consumer LLM APIs | ~$200 |
+| Funding / stress channel (Λₜ) | FRED spreads, FINRA margin, OFR indices, CBOE | public | free |
+| Cross-market L0 | TIC flows, dollar indices | public | free |
+| Alt-data layer (later) | News embeddings, positioning surveys | vendor | deferred |
+
+**The punchline stays the same as DATA_REQUIREMENTS.md:** experiments
+E1 + E2 + E4 + E5 — enough for the first real-data paper — are executable
+in one quarter by one person inside any research group with WRDS access
+plus about **$200** of API budget. The scarce resource is access, not money.
+
+### E7 — the real-data denoised-price validation (To-C line)
+
+The synthetic concept demo ([`demo/denoised_price_2026.py`](demo/denoised_price_2026.py))
+graduates to a real-data experiment:
+
+> **E7.** Reconstruct the denoised equilibrium track P^eq for the memory
+> sector through the July 2026 unwind (SOX −19%, worst month since 2008)
+> using only point-in-time data: prices (CRSP/Polygon), institutional
+> positioning (13F + COT), retail-flow proxies, and the E5 retail-AI
+> kernel. Measure: did divergence D_t cross threshold, with the
+> institutional field rotating out, materially before July 24? Deliverable:
+> a walk-forward figure exactly like the 2008 hindcast — same honesty
+> rules, no look-ahead.
+
+Data unlock: same as E1–E3 (nothing new to buy). Priority: **P0 for the
+product line** — this is the first exhibit any retail-facing partner will
+ask for.
+
+## 2 · Pillar two: training compute
+
+Phase 1 (everything in this repo) runs on a laptop; that was the point.
+Phase 2 — the [Neural Network Game Structure](docs/PHASE2_NEURAL_GAME.md),
+where every neuron is itself a small network — is where the compute bill
+arrives. Costs below are honest ranges at mid-2026 cloud prices, not
+precision estimates:
+
+| Stage | What runs | Hardware | Est. cost |
+|---|---|---|---|
+| Phase 1 (today) | HJB–FPK solvers, demos, 50 tests | laptop / free Colab | ~$0 |
+| Real-data Phase 1 (E1–E7) | Encoder training on real panel, walk-forward backtests | 1× consumer GPU or A100 spot | $1–3k |
+| Phase 2 pilots (E8–E10) | Architecture ablations on synthetic market, ~10² agent-networks | 1× H100/H200 | $5–15k |
+| Phase 2 full train (E11) | L1+L2 NNGS: ~10³ neuron-networks × 10⁵–10⁶ params, adversarial co-training + reflexivity loop | **8× H200 node, weeks-scale runs** | $100–250k per campaign |
+| Type 2 sandbox (Horizon 3+) | Agent-level "capitalism simulator" disciplined by the Type 1 equilibrium | multi-node cluster | deferred until E11 says it's earned |
+
+Why it squares: each unit's learning signal depends on every other unit's
+current policy (non-stationary co-training), and the units themselves are
+models — see the complexity-wall section of the
+[Phase 2 design doc](docs/PHASE2_NEURAL_GAME.md#5--the-complexity-wall--and-the-four-tools-against-it)
+for the four reductions (mean-field factorization, latent embeddings,
+hierarchy-as-curriculum, sparse strategic attention) that keep the bill in
+five figures for pilots rather than seven.
+
+## 3 · Use of funds, by scenario
+
+| Scenario | Budget | Buys |
+|---|---|---|
+| **Bootstrap** (status quo) | ~$500 | E5 atlas ($200) + Polygon starter + misc. Everything else free/academic. |
+| **Seed research grant** | ~$25k | All of the above + real-data Phase 1 (E1–E7) + Phase 2 pilots (E8–E10) on rented H100/H200. Output: the NeurIPS-workshop paper *and* the E7 product exhibit. |
+| **Partner / pre-seed** | ~$300k | One full NNGS training campaign (E11) + one year of data subscriptions + paper-trading infrastructure live (Airflow + Alpaca, already scaffolded in [`online/`](online/)). |
+
+No headcount is priced in: the maintainer cost of this project is one
+student who refuses to stop.
+
+---
+
+*Offering access, compute, or capital: see
+[Partnerships & Contact](README.md#partnerships--contact) — or open an
+issue.*
diff --git a/demo/denoised_price_2026.py b/demo/denoised_price_2026.py
new file mode 100644
index 0000000..fcf5e05
--- /dev/null
+++ b/demo/denoised_price_2026.py
@@ -0,0 +1,205 @@
+"""
+Denoised Equilibrium Price · 2026 — the retail product line, as a concept demo
+on SYNTHETIC data, qualitatively calibrated to the July 2026 memory-sector
+unwind (SOX −19% in July 2026, its worst month since 2008).
+
+⚠️ THIS IS A SYNTHETIC SCENARIO, NOT A HINDCAST ON REAL DATA AND NOT
+ INVESTMENT ADVICE. Every path below is generated by the dual-noise +
+ mean-field machinery of this repo. The July 2026 episode supplies the
+ *narrative shape* (institutional crowding → LLM-homogenized retail
+ chase → fundamental re-rating → forced unwind with retail capitulating
+ at the bottom while institutions re-enter); the numbers are simulated.
+ The real-data version of this demo is experiment E7 in RESOURCES.md.
+
+The product thesis (To-C line):
+ A retail investor cannot out-trade the behavioral noise ν_η — Theorem 1
+ (dual Cramér-Rao) says nobody can. But the *equilibrium* component of
+ the price — what the asset is worth once every agent has played its
+ rational strategy — is exactly what the MFG solves for. Publish that
+ denoised track and the divergence D_t = P_t/P_t^eq − 1 becomes a
+ mid/long-horizon positioning signal: when the market trades 25% above
+ its own equilibrium while the institutional mean field is already
+ rotating out, you do not need to call the day of the crash — you need
+ to not be the one buying the top.
+
+Mechanics (same objects as the rest of the repo):
+ P_t^eq — equilibrium price: the MFG-consistent fundamental track.
+ Steps down once, on the guidance shock (a Mode-I event
+ operator: genuine information, Theorem 3 irreversibility).
+ κ_t — institutional crowding (Level-1/2 mean-field concentration).
+ f_t — retail flow: momentum-chasing, LLM-homogenized (Day 16).
+ w_t — the behavioral wedge: dw = a·(κ + b·f⁺) dt − c·w dt + jumps.
+ Market price P_t = P_t^eq · (1 + w_t).
+ Signal — U_t fires when D_t > θ_D while the smoothed κ-trend has
+ turned negative for ≥3 consecutive sessions: the wedge is
+ large AND the agents holding it up are leaving.
+
+Run: python demo/denoised_price_2026.py → figures/denoised_price_demo.png
+"""
+import os
+import numpy as np
+import pandas as pd
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+FIGS = os.path.join(os.path.dirname(HERE), "figures")
+
+rng = np.random.default_rng(20260729)
+
+# ── calendar ───────────────────────────────────────────────────────────────────
+dates = pd.bdate_range(end="2026-07-29", periods=252)
+n = len(dates)
+t = np.arange(n)
+
+GUIDANCE = pd.Timestamp("2026-07-13") # HBM4-delay / guidance shock (Mode-I event)
+UNWIND = pd.Timestamp("2026-07-24") # unwind accelerates (Korea selloff spillover)
+CAPITUL = pd.Timestamp("2026-07-28") # capitulation session
+i_guid = dates.get_loc(GUIDANCE)
+i_cap = dates.get_loc(CAPITUL)
+
+# ── equilibrium (denoised) price ───────────────────────────────────────────────
+# Fundamental drift of an AI-memory franchise: strong but not manic.
+mu_eq = 0.00055
+eq_shock = np.zeros(n)
+eq_shock[i_guid:] = np.log(1 - 0.06) # −6% genuine re-rating at guidance
+log_eq = np.log(100.0) + mu_eq * t + eq_shock \
+ + np.cumsum(rng.normal(0, 0.0035, n)) # physical noise σ_τ only
+P_eq = np.exp(log_eq)
+
+# ── institutional crowding κ_t (mean-field concentration, Level 1–2) ───────────
+# Logistic build-up through the AI-memory trade; institutions begin rotating
+# out a few sessions BEFORE the guidance shock (they see the channel checks),
+# and de-crowd hard after it.
+kappa = 0.15 + 0.75 / (1 + np.exp(-(t - n * 0.55) / 18))
+i_rot = i_guid - 4
+decay = np.zeros(n)
+decay[i_rot:] = np.linspace(0, 1, n - i_rot) ** 1.35
+kappa = kappa - 0.55 * decay * kappa
+kappa = np.clip(kappa + rng.normal(0, 0.008, n), 0.05, 0.95)
+# Post-capitulation: institutions buy the dip (the AlphaGBM observation)
+kappa[i_cap:] += np.linspace(0, 0.05, n - i_cap)
+
+# ── retail flow f_t (LLM-homogenized momentum chase, Day 16) ───────────────────
+mom = pd.Series(log_eq).diff(20).fillna(0).to_numpy()
+h = 0.85 # homogenization coefficient
+f = h * np.tanh(35 * mom) + rng.normal(0, 0.06, n)
+f[i_guid:i_cap] += 0.25 # retail buys the first dip…
+f[i_cap:] = -0.9 + rng.normal(0, 0.05, n - i_cap) # …then capitulates at the low
+
+# ── behavioral wedge w_t ───────────────────────────────────────────────────────
+w = np.zeros(n)
+a, b, c = 0.0030, 0.55, 0.0105
+for k in range(1, n):
+ drive = a * (kappa[k] + b * max(f[k], 0.0))
+ jump = 0.0
+ if rng.uniform() < 0.04 + 0.10 * w[k - 1]: # jumps cluster with crowding
+ jump = rng.choice([-1, 1], p=[0.35, 0.65]) * rng.exponential(0.006)
+ if k >= i_cap - 2: # forced unwind: 4 sessions
+ drive, c_eff = -0.055, 0.28
+ jump = -abs(rng.normal(0.012, 0.006))
+ else:
+ c_eff = c
+ w[k] = w[k - 1] + drive - c_eff * w[k - 1] + jump
+w = np.clip(w, -0.08, 0.40)
+
+P = P_eq * (1 + w)
+D = P / P_eq - 1 # divergence
+
+# ── the signal U_t ─────────────────────────────────────────────────────────────
+THETA_D = 0.20
+kap_trend = pd.Series(kappa).ewm(span=10).mean().diff()
+falling = (kap_trend < 0).rolling(3).sum() == 3
+armed = (pd.Series(D) > THETA_D) & falling.fillna(False)
+sig_idx = armed[armed & (armed.index >= i_guid - 10)].index.min()
+sig_date = dates[sig_idx]
+lead_unwind = int(((dates > sig_date) & (dates <= UNWIND)).sum())
+lead = int(((dates > sig_date) & (dates <= CAPITUL)).sum())
+
+peak = P.max()
+trough = P[i_cap - 2:].min()
+
+print("=" * 70)
+print("SYNTHETIC CONCEPT DEMO — denoised equilibrium price (To-C product line)")
+print(f"scenario calendar : {dates[0].date()} → {dates[-1].date()} ({n} sessions)")
+print(f"guidance shock : {GUIDANCE.date()} (Mode-I operator, eq −6%)")
+print(f"signal U_t fires : {sig_date.date()} — D = {D[sig_idx]*100:+.1f}% above "
+ f"equilibrium, κ-trend negative 3 sessions")
+print(f"lead time : {lead_unwind} trading days (≈2 weeks) before the "
+ f"{UNWIND.date()} unwind acceleration,")
+print(f" {lead} trading days before the {CAPITUL.date()} capitulation")
+print(f"peak → trough : {(trough/peak - 1)*100:.0f}% "
+ f"(market returns to its denoised track)")
+print(f"at the trough : retail flow f = {f[i_cap+1]:+.2f} (panic sell) | "
+ f"institutional κ rising (buying the dip)")
+print("=" * 70)
+
+# ── figure ─────────────────────────────────────────────────────────────────────
+OK = dict(blue="#0072B2", orange="#E69F00", red="#D55E00", ink="#111827",
+ mute="#6B7280", grid="#E5E7EB")
+plt.rcParams.update({"font.family": "sans-serif", "font.size": 10,
+ "axes.spines.top": False, "axes.spines.right": False,
+ "figure.facecolor": "white"})
+fig, axes = plt.subplots(2, 1, figsize=(13.2, 7.6), sharex=True,
+ gridspec_kw=dict(height_ratios=[1.15, 1.0], hspace=0.10))
+lo, hi = dates[0], dates[-1]
+
+ax = axes[0]
+ax.plot(dates, P, color=OK["ink"], lw=1.5, label="market price $P_t$ (equilibrium + behavioral wedge)")
+ax.plot(dates, P_eq, color=OK["blue"], lw=1.7,
+ label="denoised equilibrium price $P_t^{eq}$ (the product)")
+ax.fill_between(dates, P_eq, P, where=P > P_eq, color=OK["orange"], alpha=0.25,
+ label="behavioral wedge $w_t$ (crowding, Thm 1)")
+ax.axvline(sig_date, color=OK["orange"], lw=1.5, ls="--")
+ax.axvline(UNWIND, color=OK["red"], lw=1.0, ls=":")
+ax.axvline(CAPITUL, color=OK["red"], lw=1.4, ls="--")
+ax.annotate(f"{sig_date.strftime('%b %d')} — signal: price {D[sig_idx]*100:+.0f}% above\n"
+ f"equilibrium while institutions rotate out.\n"
+ f"{lead_unwind} trading days (≈2 weeks) before the\n"
+ f"unwind accelerates on Jul 24.",
+ xy=(sig_date, P[sig_idx]), xytext=(dates[int(n*0.35)], peak * 0.915),
+ fontsize=9.5, fontweight="bold", color="#B45309",
+ arrowprops=dict(arrowstyle="->", color="#B45309", lw=1.2))
+ax.annotate("Jul 28 — capitulation:\nretail panic-sells at the low,\n"
+ "institutions buy; price re-joins\nits denoised track.",
+ xy=(CAPITUL, trough), xytext=(dates[int(n*0.46)], P_eq[0] * 0.985),
+ fontsize=9.5, fontweight="bold", color=OK["red"],
+ arrowprops=dict(arrowstyle="->", color=OK["red"], lw=1.2,
+ connectionstyle="arc3,rad=0.15"))
+ax.axvline(GUIDANCE, color=OK["mute"], lw=0.9, ls=":")
+ax.text(GUIDANCE - pd.Timedelta(days=3), 119.5, "guidance shock\nJul 13 (Mode-I)",
+ fontsize=7.8, color=OK["mute"], ha="right")
+ax.set_ylabel("price (indexed)")
+ax.legend(loc="upper left", fontsize=8.5, frameon=False)
+ax.set_title("Denoised equilibrium price — SYNTHETIC concept demo of the To-C product line "
+ "(scenario shaped on the July 2026 memory unwind)",
+ fontsize=12.5, fontweight="bold", loc="left", pad=10)
+ax.grid(alpha=0.3)
+
+ax = axes[1]
+ax.plot(dates, D * 100, color=OK["ink"], lw=1.5, label="divergence $D_t = P_t/P_t^{eq}-1$")
+ax.axhline(THETA_D * 100, color=OK["red"], ls=":", lw=1.3)
+ax.text(dates[int(n*0.42)], THETA_D * 100 + 1.2, f"θ_D = {THETA_D*100:.0f}%",
+ fontsize=8.5, color=OK["red"])
+ax.fill_between(dates, THETA_D * 100, np.clip(D * 100, THETA_D * 100, None),
+ color=OK["orange"], alpha=0.45)
+ax2 = ax.twinx()
+ax2.plot(dates, kappa, color=OK["blue"], lw=1.3, alpha=0.85)
+ax2.set_ylabel("institutional crowding κ_t", color=OK["blue"])
+ax2.tick_params(axis="y", labelcolor=OK["blue"])
+ax2.spines.top.set_visible(False)
+ax.axvline(sig_date, color=OK["orange"], lw=1.5, ls="--")
+ax.axvline(CAPITUL, color=OK["red"], lw=1.4, ls="--")
+ax.set_ylabel("divergence (%)")
+ax.set_xlim(lo, hi)
+ax.legend(loc="upper left", fontsize=8.5, frameon=False)
+ax.grid(alpha=0.3)
+
+fig.text(0.01, 0.005,
+ "SYNTHETIC DATA — a concept demo of the denoised-price product, not investment advice and not "
+ "a real-data hindcast. Signal U_t: D_t > θ_D while the EMA(10) κ-trend is negative 3 consecutive "
+ "sessions. Real-data version: experiment E7 (RESOURCES.md).",
+ fontsize=7.5, color=OK["mute"])
+fig.savefig(os.path.join(FIGS, "denoised_price_demo.png"), dpi=200, bbox_inches="tight")
+print("✓ figures/denoised_price_demo.png")
diff --git a/docs/JOURNEY.md b/docs/JOURNEY.md
new file mode 100644
index 0000000..02ee41f
--- /dev/null
+++ b/docs/JOURNEY.md
@@ -0,0 +1,155 @@
+# The Journey — From World-Model Skeptic to This Repository
+
+*HongJin HE · Alpha Flow Research · July 2026*
+
+*[README](../README.md) · [中文](../README_CN.md)*
+
+---
+
+## I did not believe in world models
+
+For a long time I thought "world model" was a marketing term, and I had two
+objections that felt unanswerable.
+
+**The precision objection.** Quantum mechanics puts a hard floor under how
+precisely the state of a physical system can be known — and chaotic dynamics
+amplify any microscopic uncertainty exponentially. If you cannot pin down the
+state, how can you claim to model the world that generates it?
+
+**The storage objection.** Even granting perfect knowledge, a faithful
+simulation of the world explodes combinatorially. Every naive estimate of
+"simulate the environment" lands orders of magnitude beyond any hardware
+roadmap. A world model, I concluded, was either a toy or a fantasy.
+
+I was wrong about both — but it took watching three fields succeed in
+parallel to see why.
+
+## What changed my mind
+
+**Robots that dream.** The lineage that started with Ha & Schmidhuber's
+[World Models](https://worldmodels.github.io/) (2018) — an agent learning
+inside its own compressed dream of the environment — became Hafner's
+[Dreamer](https://arxiv.org/abs/1912.01603) line, and by
+[DreamerV3](https://github.com/danijar/dreamerv3) (2023) a single
+configuration was mastering 150+ domains from pixels. The result that
+genuinely shook me was
+[DayDreamer](https://danijar.com/project/daydreamer/) (Wu, Escontrela,
+Hafner, Abbeel & Goldberg, CoRL 2022): a *physical* quadruped learning to
+walk in about one hour, because it practiced inside a learned latent model
+instead of on its own legs. No atom of the robot's world was simulated. Only
+what the task needed was kept.
+
+**Cars that predict.** Wayve's [GAIA-1](https://arxiv.org/abs/2309.17080)
+(2023) and [GAIA-2](https://wayve.ai/thinking/gaia-2/) (2025) generate
+coherent futures of driving scenes — not because they track every photon on
+the road, but because they learned the *distribution of plausible futures*
+at exactly the abstraction level where driving decisions live. Autonomous
+driving stopped asking "what is the exact state of the world?" and started
+asking "what happens next, at the resolution that matters?"
+
+**Agents that imagine.** DeepMind's [Genie](https://arxiv.org/abs/2402.15391)
+(ICML 2024) through [Genie 3](https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/)
+(2025) learn playable, interactive worlds from video; agents make decisions
+by *imagining* rollouts. [NVIDIA Cosmos](https://github.com/NVIDIA/Cosmos)
+(2025) industrialized the recipe into world foundation models for physical
+AI. And Yann LeCun's position paper,
+[A Path Towards Autonomous Machine Intelligence](https://openreview.net/forum?id=BZ5a1r-kVsf)
+(2022), together with Meta's [V-JEPA 2](https://arxiv.org/abs/2506.09985)
+(2025), articulated the principle underneath all of it: **predict in
+representation space, not in pixel space**. You do not model the world.
+You model the *sufficient statistics* of the world for the decisions you
+need to make.
+
+That principle dissolved both of my objections at once:
+
+- The **storage objection** dies because compression is the whole point.
+ A world model is not a simulation of the world; it is the smallest state
+ that makes the future predictable. (In this repo, that state is 5
+ dimensions per asset plus a distribution — not a tick-by-tick replay.)
+- The **precision objection** dies because prediction never needed
+ microscopic precision. Boltzmann could not track a single molecule, and
+ kinetic theory works anyway: at the population level, dynamics become
+ *more* lawful as N grows, not less. Quantum indeterminacy lives twenty
+ orders of magnitude below the level where any decision — a lane change, a
+ portfolio weight — actually happens.
+
+I was an exchange student at Stanford while much of this was in the air —
+the [spatial intelligence](https://www.ted.com/talks/fei_fei_li_with_spatial_intelligence_ai_will_understand_the_real_world)
+conversation around Fei-Fei Li and [World Labs](https://www.worldlabs.ai/)
+made it feel less like a research direction and more like a consensus
+forming in real time: *world models are how agents will understand
+everything*. The question I could not put down was: **everything — except
+markets?**
+
+## The second thread: a violation of first principles
+
+At the same time I was working through quantitative finance coursework and
+reading the empirical asset-pricing literature, and I hit something I still
+find astonishing: **the dominant paradigm of quantitative finance is not
+even trying to predict the market.**
+
+Factor models ask which characteristics *correlate* with cross-sectional
+returns. The literature documented 600+ of them — Cochrane called it the
+["factor zoo"](https://onlinelibrary.wiley.com/doi/10.1111/j.1540-6261.2011.01671.x)
+in his 2011 AFA presidential address — and when
+[Harvey, Liu & Zhu](https://academic.oup.com/rfs/article/29/1/5/1843824)
+audited the zoo, most factors failed to replicate.
+[López de Prado](https://www.wiley.com/en-us/Advances+in+Financial+Machine+Learning-p-9781119482086)
+catalogued the same pathology from inside the industry: backtest overfitting
+as standard practice.
+
+From first principles this is upside down. A model of the market should
+model *the market* — the thing that generates prices — not mine correlations
+from the residue prices leave behind. History's data does not repeat: every
+factor decays the moment it is crowded (the Lucas critique, measured).
+But history itself — the *structure* that generates the data: institutions,
+constraints, incentives, the game — does repeat. If you want factors that
+mean something, you have to go back to the world that produces them.
+
+That sentence, I realized, had a name. It was a world model.
+
+## The merge
+
+Two threads, one collision:
+
+- Robotics, driving, and agent research proved that **compressed world
+ models predict well** at the abstraction level that matters.
+- Quantitative finance was stuck mining patterns precisely because it had
+ **no world model** — no representation of the thing generating the data.
+
+But markets add one twist that Dreamer never faced: **the "physics" of a
+market is other agents' strategies.** A road does not replan when GAIA-1
+predicts it; a market does — every fund that discovers a pattern destroys
+it by trading on it. So the transition network of a financial world model
+cannot be learned dynamics; it must be an **equilibrium solver**. The
+environment *is* the fixed point of every agent's best response to everyone
+else.
+
+That single substitution — replace the learned transition model with a
+hierarchical mean-field game — is this entire repository. The encoder
+compresses the market panel into a latent state (the JEPA lesson). The Game
+module solves for the equilibrium instead of extrapolating history (the
+Lucas lesson). The controller acts on the equilibrium drift (the Dreamer
+lesson). And the residual that no data can remove — behavioral noise — is
+not swept under the carpet but bounded by theorem
+([Theorem 1](../README.md#component-2--dual-noise-decomposition-theorem-1),
+the dual Cramér-Rao bound): the honest descendant of my old precision
+objection.
+
+## What I believe now
+
+World models are not a technique that happens to work in robotics. They are
+the first-principles form of prediction in any domain: find the level of
+description at which dynamics are lawful, compress to it, and model the
+generator — not the residue it leaves in a dataset. Finance was simply the
+domain where nobody had done it yet, because in finance the generator
+fights back.
+
+MicroWorld is the bet that modeling the generator anyway — as a game, with
+proofs — is worth more than one more decade of factor mining.
+
+---
+
+*Back to the [README](../README.md) · the mathematics starts
+[here](../README.md#the-mathematical-framework-two-threads-one-theory) ·
+what comes after the mathematics is [Phase 2](PHASE2_NEURAL_GAME.md).*
diff --git a/docs/ONE_PAGER.md b/docs/ONE_PAGER.md
new file mode 100644
index 0000000..38410eb
--- /dev/null
+++ b/docs/ONE_PAGER.md
@@ -0,0 +1,70 @@
+# MicroWorld — One Pager
+
+**A world model for equity markets: model the players, not the patterns.**
+*Alpha Flow Research · HongJin HE · HKUST / Stanford IHP · July 2026*
+*[github.com/hongjin-he/MicroWorld](https://github.com/hongjin-he/MicroWorld) · MIT license*
+
+---
+
+**The problem.** Quantitative finance's dominant paradigm — factor mining
+and time-series ML — extracts patterns from historical data. But historical
+*data* does not repeat: every discovered signal is destroyed by its own
+adoption (alpha half-life: ~6 years in 1990 → ~11 months in 2023). What
+does repeat is the *structure* that generates the data: institutions,
+regulations, incentives, the game.
+
+**The approach.** MicroWorld models that structure directly: US equity
+markets as a four-level hierarchical mean-field game — cross-market flows,
+institution types, individual institutions, intra-institution desks — with
+a 5-dimensional state space per asset, a dual decomposition of noise into
+physical and behavioral components, and a 22-operator event algebra for
+M&A/IPO/policy shocks. Predictions are equilibria, not extrapolations, so
+they are designed to survive their own deployment. Seven theorems proved;
+50 tests passing; every demo runs on a laptop with zero API keys.
+
+**Validation to date.**
+- **2008 hindcast, public data only:** the stability indicator Λₜ entered
+ its sustained crisis regime on Aug 16, 2007 — **272 trading days before
+ Lehman** — with zero false alarms in 2005–06. Same signal: Feb 20, 2020.
+- **Synthetic product demo:** denoised equilibrium price flags a
+ crowding-driven divergence ~2 weeks before a July-2026-shaped sector
+ unwind ([demo](../demo/denoised_price_2026.py)); real-data version is
+ specified as experiment E7.
+- **Open-source traction:** 138+ GitHub stars in the first weeks, 17-notebook
+ tutorial series, interactive 3D market universe.
+
+**Two products, one engine.**
+- **To-C — the denoised price.** The MFG equilibrium track P^eq and the
+ divergence D_t = P/P^eq − 1: a mid/long-horizon positioning research
+ signal for retail ("are you buying value or buying crowding?"). Not
+ personalized advice; an instrument-level research layer.
+- **To-B — structural risk early warning.** Λₜ regime monitoring, crowding
+ decomposition, and event-operator scenario analysis for funds and risk
+ desks — the 2008-grade signal, live (pipeline scaffolded: Airflow +
+ Alpaca paper-trading).
+
+**The moat.** Causal explainability. Factor and time-series shops find
+patterns they cannot explain and therefore cannot defend when regimes
+break. Every MicroWorld output is attributable to agents, constraints, and
+equilibrium conditions — the model's explanation *is* the model. The
+framework is also the only one simultaneously offering strategic agents,
+universe-changing events, noise decomposition with an estimation bound,
+crisis early-warning, and a four-level hierarchy.
+
+**Roadmap.** Phase 1 (now): real-data validation E1–E7 → NeurIPS-2026
+workshop paper + the E7 product exhibit (~one quarter, ~$25k inc. compute).
+Phase 2: the Neural Network Game Structure — every neuron an agent-network,
+regulatory constraints as architecture ([design doc](PHASE2_NEURAL_GAME.md));
+pilots on one H100/H200, full campaign on an 8×H200 node (~$300k scenario).
+Detail: [RESOURCES.md](../RESOURCES.md).
+
+**The ask.** Data access (WRDS-grade), compute (H100/H200 hours), or
+pre-seed partnership — priced by scenario in RESOURCES.md.
+
+**Contact.** [LinkedIn](https://www.linkedin.com/in/hongjinhe-hkust-edu) ·
+[GitHub](https://github.com/hongjin-he) · [X](https://x.com/Mr_Abstractor)
+
+---
+
+*This document describes research software. Nothing here is investment
+advice or an offer of securities.*
diff --git a/docs/PHASE2_NEURAL_GAME.md b/docs/PHASE2_NEURAL_GAME.md
new file mode 100644
index 0000000..a46229c
--- /dev/null
+++ b/docs/PHASE2_NEURAL_GAME.md
@@ -0,0 +1,171 @@
+# Phase 2 — The Neural Network Game Structure (NNGS)
+
+**Status: design document.** Nothing in this file is implemented, and that
+is deliberate. Phase 2 begins when two resources exist that do not exist
+today: the cleaned multi-type data panel and H200-class training compute
+(see [RESOURCES.md](../RESOURCES.md)). Until then, this document is the
+specification we hold ourselves to — written down in public, before the
+hardware arrives, so the idea has a timestamp.
+
+*[README](../README.md) · [The journey that led here](JOURNEY.md)*
+
+---
+
+## 1 · Why Phase 1 is a game-theory engine, not a neural network
+
+Phase 1 — everything currently in this repository — solves the market as a
+hierarchical mean-field game: coupled HJB–FPK systems, an operator algebra
+for events, seven proven theorems. It is deliberately *mathematical
+structure first, learning second*, for one reason: **data scarcity is not a
+temporary inconvenience; at today's access level it is binding.** When you
+cannot estimate millions of parameters, you must get the same behavior from
+structure — and Theorem 1 (the dual Cramér-Rao bound) says something
+stronger: part of the residual can *never* be estimated away with more
+data, only explained by a model of the mechanism. Phase 1 is the
+mechanism, written in PDEs because PDEs are what run on a laptop.
+
+So the honest description of the current stage is: **the game-theoretic
+core is the data-efficient regime of the world model.** It is not the
+final form.
+
+## 2 · The Phase 2 thesis: neurons that play, not neurons that fire
+
+Everyone else who models markets builds a mathematical model *of* the
+agents. Phase 2 builds the agents *as the network*:
+
+> **A neural network in which each neuron is itself a small neural
+> network — one per institution or retail cohort — with its own objective,
+> its own information set, and its own strategy. The connections between
+> neurons are not weights that passively mix signals; they are strategic
+> couplings: each unit's forward pass is a best response to the others.
+> We call this the Neural Network Game Structure (NNGS).**
+
+The units are, in a precise sense, *personified*: a neuron here does not
+"fire", it *decides*. Where a GNN node aggregates its neighbors' messages,
+an NNGS node responds to its neighbors' strategies — the difference between
+diffusion and game play. This is the same distinction that separates this
+repo from swarm models (see the MicroFish comparison in the README), now
+pushed down into the architecture itself.
+
+### Two representational choices, both admissible
+
+1. **Agents-as-neurons (the primary design).** Each of the ~11 agent
+ classes (6 institutional + 5 retail, [`agents/`](../agents/)) is
+ instantiated as a population of small networks, fully connected across
+ the graph, each carrying its own recurrent state.
+2. **Levels-as-layers (the fallback).** The four levels L0–L3 become four
+ layers of one large network, with within-layer competition expressed via
+ lateral connections. Coarser, cheaper, less faithful — but trainable
+ sooner.
+
+These are two encodings of the same object; which one wins is an empirical
+question (experiment **E9** below), not an aesthetic one.
+
+## 3 · The constraint structure — why this is not a free-form black box
+
+An unconstrained network-of-networks would forfeit the one thing Phase 1
+paid for: interpretability. NNGS keeps it through **constraints that are
+facts about the world, imposed as architecture**:
+
+- **Regulatory weight sharing.** All agents of one type live under the same
+ regulator — every bank under Basel III, every mutual fund under UCITS/40-Act
+ limits, every insurer under Solvency-style capital rules. Architecturally:
+ agents of a type share a constraint module (projection layer onto the
+ feasible set), exactly as the Phase 1 taxonomy shares Merton-style
+ constraint sets. The shared module *is* the regulator.
+- **Budget and leverage feasibility** as hard projection layers, not soft
+ penalties — an agent cannot learn its way out of a balance sheet.
+- **Information stratification** ([`state/information.py`](../state/information.py)):
+ each neuron sees only its type's filtration. A retail cohort cannot
+ attend to order-flow features it would not observe in reality.
+- **The mean-field anchor.** The Phase 1 equilibrium is retained as a
+ regularizer: population-level statistics of NNGS play must stay within a
+ Wasserstein ball of the MFG equilibrium μ*, unless data demands
+ otherwise. Phase 1 becomes the prior; Phase 2 the posterior.
+
+The result is a *constrained* neural network whose every departure from
+equilibrium is attributable — to an agent type, a constraint, or an
+information set. Explanations survive the scaling-up.
+
+## 4 · One network or two? — an experiment, not a debate
+
+Two candidate macro-architectures:
+
+| | **A — Unified** | **B — Dual (environment + population)** |
+|---|---|---|
+| Structure | One network: agents and environment dynamics entangled in a single graph | Network 1: a trained environment/world model (the neuralized **E**). Network 2: the population of agent-neurons playing inside it (the neuralized **Game**) |
+| Lineage | End-to-end world models (Dreamer-style) | E-Game-C itself — B is its native neuralization |
+| Risk | Attribution becomes murky; environment leaks into strategy | Interface mismatch: the population's actions must feed back into the environment model consistently (reflexivity must close the loop) |
+| Prior | — | Phase 1 MFG initializes Network 2; the Phase 1 encoder initializes Network 1 |
+
+Design B is the default because it inherits E-Game-C directly and keeps the
+reflexivity loop explicit (price-belief feedback as the interface between
+the two networks). But the choice is assigned to experiment **E8**, run
+first at toy scale — the answer we publish will be measured, not asserted.
+
+## 5 · The complexity wall — and the four tools against it
+
+Honesty first: training a model whose parameters are themselves models
+roughly **squares the training difficulty**. N agents with pairwise
+strategic coupling is O(N²) in interactions before anything recurses, and
+each unit's learning signal depends on every other unit's current policy —
+the non-stationarity that makes multi-agent RL notoriously unstable, at a
+scale multi-agent RL has not attempted. This is exactly why Phase 2 waits
+for hardware, and why the design leans on four reductions:
+
+1. **Mean-field factorization.** Within a type, agents couple to the
+ *distribution* of their type, not to each individual — O(N²) → O(N·K)
+ for K types. This is not an approximation bolted on: Phase 1 *proved*
+ (Theorem 7.4, propagation of chaos) that it is the correct N→∞ limit.
+2. **Latent agent embeddings.** Each neuron's policy conditions on a
+ low-dimensional embedding of its identity (d ≈ 16–64), so populations
+ share one policy network modulated per agent — the DreamerV3 trick of
+ one configuration spanning many domains, applied within one market.
+3. **Hierarchy as curriculum.** Train L1 (type-level) frozen-environment
+ first, unfreeze L2 (institution-level), then L3 — the four-level
+ structure is not just descriptive, it is the training schedule.
+4. **Sparse strategic attention.** Full connectivity is the specification,
+ not the runtime: a learned top-k attention over counterparties captures
+ the empirically sparse strategic graph (a fund responds to its actual
+ competitors, not to all 50,000 institutions).
+
+## 6 · Compute and data triggers
+
+- **Compute.** The E8/E9 toy-scale ablations run on a single H100/H200.
+ A credible full-scale L1+L2 NNGS train (≈10³ neuron-networks, each
+ 10⁵–10⁶ parameters, adversarial co-training) is an H200-cluster problem
+ — the training-run budget lives in [RESOURCES.md](../RESOURCES.md).
+- **Data.** The cleaned, typed panel of DATA_REQUIREMENTS.md (13F/COT
+ positioning for institutional ground truth, the E5 LLM-query atlas for
+ retail policy priors). NNGS without agent-level ground truth would be a
+ simulator, not a world model.
+
+## 7 · Phase 2 experiments
+
+Continuing the E-numbering from [DATA_REQUIREMENTS.md](../DATA_REQUIREMENTS.md)
+(E1–E6) and [RESOURCES.md](../RESOURCES.md) (E7):
+
+| ID | Experiment | Question it settles | Scale |
+|---|---|---|---|
+| **E8** | Unified vs dual architecture on the synthetic market ([`demo/synthetic_market.py`](../demo/synthetic_market.py)) | One network or two? | 1 GPU |
+| **E9** | Agents-as-neurons vs levels-as-layers, same data, same budget | Which encoding of the game? | 1 GPU |
+| **E10** | MFG-distillation: initialize NNGS from the Phase 1 equilibrium vs cold start | Is Phase 1 a useful prior (we predict: decisively yes)? | 1 GPU → cluster |
+| **E11** | Constrained vs unconstrained NNGS on real panel | Do the regulatory constraints help or hurt fit? (The thesis: they *are* the alpha) | cluster |
+
+## 8 · The larger claim
+
+If NNGS works for markets, nothing about it is market-specific. Any
+real-world graph whose nodes have objectives — supply chains, electricity
+markets, ecosystems of platforms, geopolitical blocs — admits the same
+construction: **model each node as a small neural network with a stake,
+and the graph as their game.** Markets are simply the best first target:
+the players are catalogued, the constraints are written law, and the
+scoreboard prints every millisecond.
+
+That is the Phase 2 bet. The mathematics of Phase 1 is how we earn the
+right to place it.
+
+---
+
+*Questions, objections, or compute to offer: see
+[Partnerships](../README.md#partnerships--contact).*
diff --git a/docs/i18n/README_JA.md b/docs/i18n/README_JA.md
new file mode 100644
index 0000000..6440f54
--- /dev/null
+++ b/docs/i18n/README_JA.md
@@ -0,0 +1,104 @@
+
+
+# MicroWorld
+# 世界モデル × クオンツファイナンス
+
+*金融市場のために設計された初の世界モデル・アーキテクチャ*
+*従来のファクター・マイニングを超える、クオンツファイナンスの新パラダイム*
+
+**[English](../../README.md) | [中文文档](../../README_CN.md) | [日本語](README_JA.md) | [한국어](README_KO.md)**
+
+**Alpha Flow Research · HongJin HE · HKUST / Stanford IHP · 2026年7月**
+
+
+
+> これは完全版READMEの要約翻訳です。数学的詳細・17日間ノートブック・全実験計画は[英語版](../../README.md)をご覧ください。
+
+---
+
+## 2008年9月12日
+
+> *リーマン・ブラザーズが破綻申請する直前の金曜日、VIXの終値は25.7——高めだが、平常の範囲。その13ヶ月前、本モデルの危機指標はすでに安定レジームを離脱していた。*
+
+[](../../demo/hindcast_2008.py)
+
+**公開データのみによる2005–2009年のウォークフォワード再現**——各時点で「その朝までに入手可能な情報」だけを使用：
+
+- 2005–2006年：誤警報**ゼロ**
+- **2007年8月16日**——Λₜが持続的危機レジーム入り：**リーマン破綻の272営業日前**。このシグナルから2009年3月の底まで、市場は−49%
+- これはヒンドキャスト（先読みなしのバックテスト）であり、ライブ予測ではありません。教訓は「暴落の日付を当てられる」ことではなく、**安定性からの離脱は、価格ベースの指標が「平常」を示す間に、状態空間の幾何学には1年前から見えていた**ということ
+
+## プロジェクトの物語
+
+なぜすべてのトレーディングモデルは、本番投入の瞬間から死に始めるのか？答えはより良いデータでも、より大きなネットワークでもない。**市場はデータセットではなく、ゲームである。** パターンを予測すれば、パターンは消える。モデル化できる残りのものは、プレイヤー自身だけだ。
+
+MicroWorldは、米国株式市場を**4層の階層的平均場ゲーム**としてモデル化する、初のオープンソース世界モデル・アーキテクチャ：
+
+- **Level 0** — 市場間資本フロー
+- **Level 1** — 機関タイプ間の競争（中央銀行・銀行・クオンツファンド・個人投資家）
+- **Level 2** — タイプ内の個別機関
+- **Level 3** — 機関内部のデスクと個人
+
+7つの定理を証明済み。50のテストが通過。すべてのデモはAPIキーなしでラップトップ上で動作。
+
+## 第一原理の賭け——相関ではなく因果
+
+主流のクオンツ手法——ファクター・マイニング、時系列ML、特徴量エンジニアリング——はすべて、「**歴史のデータは繰り返す**」という未検証の前提の上に立つ。だが繰り返さない。発見されたパターンは、発見されたがゆえに消える（ルーカス批判）。
+
+本当に繰り返すのはデータではなく、**データを生成する世界**——マンデートを持つ機関、ルールを持つ規制当局、シグナルが公表されても変わらないインセンティブ構造だ。だから本プロジェクトは、データの残渣ではなく生成器そのものをモデル化する：
+
+> **私たちはデータの中に構造を見つけたのではない。データを生み出す構造をモデル化したのだ。**
+
+帰結は二つ：(1) すべての出力は、名前を持つ主体・制約・均衡条件に帰属可能——説明可能性は後付け機能ではなく推論の方向そのもの。(2) 均衡である予測は、実行されても蒸発しない——実行されることこそ、均衡が自己を実現する方法である。
+
+## 同一エンジン、二つのプロダクト
+
+### To-C · ノイズ除去済み均衡価格（中長期保有向け）
+
+定理1（双対クラメール・ラオ限界）：行動ノイズν_ηは誰にも取引で打ち負かせない。長期投資家が本当に必要とするのは、ノイズの下にある**均衡トラックP^eq**と乖離度**D_t = P_t/P^eq_t − 1**——「あなたが買っているのは価値か、それとも過密か？」
+
+[](../../demo/denoised_price_2026.py)
+
+上図は**合成データ**によるコンセプトデモ（2026年7月のメモリ半導体セクター崩落——SOX指数が単月−19%、2008年以来最悪——を情景の原型とする）。シミュレーション内では、シグナルが崩落加速の**11営業日（約2週間）前**に点灯。実データ版は実験E7として仕様化済み。
+
+> *研究シグナルであり、投資助言ではありません。*
+
+### To-B · 機関向け構造リスク
+
+Λₜレジーム監視（2008/2020級のシグナル）、ポジション別の過密度分解、イベント演算子によるシナリオ分析。Airflow + Alpacaのパイプラインはスキャフォールド済み。
+
+## ロードマップ——第1段階と第2段階
+
+**第1段階（現在）**：ゲーム理論コア。HJB–FPK連立系、22のイベント演算子、NeurIPSワークショップ向け実験E1–E7。データが希少な時代の理論的に正しい選択——定理1は、残差の一部が機構によって*しか*説明できないことを証明している。
+
+**第2段階：[ニューラルネットワーク・ゲーム構造（NNGS）](../PHASE2_NEURAL_GAME.md)**——**各ニューロン自体が小さなニューラルネットワーク**（機関・個人投資家コホートごとに一つ）であり、規制の現実がアーキテクチャとして課され、順伝播そのものがゲームのプレイである。訓練難度はおよそ二乗化するため、H200級の計算資源を待つ——設計文書は今、公開する。
+
+## クイックスタート
+
+```bash
+git clone https://github.com/hongjin-he/MicroWorld
+cd MicroWorld
+pip install -r requirements.txt
+
+python demo/global_demo.py # アニメーション世界モデルデモ（約2分、CPU）
+python demo/hindcast_2008.py # 2008年ヒンドキャスト
+python demo/denoised_price_2026.py # ノイズ除去価格デモ（合成データ）
+python -m pytest tests/ -v # テストスイート（50件）
+```
+
+## リンク
+
+- **完全版README**：[English](../../README.md) · [中文](../../README_CN.md)
+- **数学論文**（全証明・25頁）：[companion paper](https://github.com/hongjin-he/mathmatical-framework-for-world-models-in-quant-finance)
+- **開発の心路歴程**：[docs/JOURNEY.md](../JOURNEY.md)
+- **提携・連絡**：[LinkedIn](https://www.linkedin.com/in/hongjinhe-hkust-edu) · [X](https://x.com/Mr_Abstractor) · [GitHub issues](https://github.com/hongjin-he/MicroWorld/issues)
+
+> *コンプライアンス注記：本リポジトリは研究ソフトウェアです。いかなる内容も投資助言ではありません。*
+
+
+
+**Alpha Flow Research · HKUST · Stanford IHP · 2026年7月**
+
+*これはクオンツツールではない。金融市場を理解するための新しいパラダイムである。*
+
+
diff --git a/docs/i18n/README_KO.md b/docs/i18n/README_KO.md
new file mode 100644
index 0000000..a212c36
--- /dev/null
+++ b/docs/i18n/README_KO.md
@@ -0,0 +1,104 @@
+
+
+# MicroWorld
+# 월드 모델 × 퀀트 금융
+
+*금융 시장을 위해 설계된 최초의 월드 모델 아키텍처*
+*전통적 팩터 마이닝을 넘어서는 퀀트 금융의 새로운 패러다임*
+
+**[English](../../README.md) | [中文文档](../../README_CN.md) | [日本語](README_JA.md) | [한국어](README_KO.md)**
+
+**Alpha Flow Research · HongJin HE · HKUST / Stanford IHP · 2026년 7월**
+
+
+
+> 이 문서는 전체 README의 요약 번역본입니다. 수학적 세부 사항, 17일 노트북 시리즈, 전체 실험 계획은 [영어판](../../README.md)을 참조하세요.
+
+---
+
+## 2008년 9월 12일
+
+> *리먼 브라더스 파산 신청 직전 금요일, VIX 종가는 25.7 — 다소 높지만 평범한 수준. 그러나 13개월 전, 이 모델의 위기 지표는 이미 안정 레짐을 이탈해 있었다.*
+
+[](../../demo/hindcast_2008.py)
+
+**공개 데이터만으로 수행한 2005–2009 워크포워드 재현** — 매 시점에서 "그날 아침까지 입수 가능한 정보"만 사용:
+
+- 2005–2006년: 오경보 **0건**
+- **2007년 8월 16일** — Λₜ가 지속적 위기 레짐 진입: **리먼 파산 272거래일 전**. 이 신호부터 2009년 3월 저점까지 시장은 −49%
+- 이것은 힌드캐스트(미래 참조 없는 백테스트)이며 실시간 예측이 아닙니다. 교훈은 "폭락 날짜를 맞출 수 있다"가 아니라, **가격 기반 지표가 "평상시"를 가리키는 동안, 안정성 이탈은 상태 공간의 기하학에는 1년 전부터 보였다**는 것
+
+## 프로젝트 이야기
+
+왜 모든 트레이딩 모델은 실전 투입 순간부터 죽기 시작하는가? 답은 더 좋은 데이터도, 더 큰 네트워크도 아니다. **시장은 데이터셋이 아니라 게임이다.** 패턴을 예측하면 패턴은 사라진다. 모델링할 수 있는 유일한 대상은 플레이어 자신이다.
+
+MicroWorld는 미국 주식 시장을 **4계층 위계적 평균장 게임(mean-field game)**으로 모델링하는 최초의 오픈소스 월드 모델 아키텍처:
+
+- **Level 0** — 시장 간 자본 흐름
+- **Level 1** — 기관 유형 간 경쟁 (중앙은행 · 은행 · 퀀트 펀드 · 개인투자자)
+- **Level 2** — 유형 내 개별 기관
+- **Level 3** — 기관 내부의 데스크와 개인
+
+7개 정리 증명 완료. 50개 테스트 통과. 모든 데모는 API 키 없이 노트북에서 실행.
+
+## 제1원리의 베팅 — 상관이 아닌 인과
+
+주류 퀀트 방법론 — 팩터 마이닝, 시계열 ML, 피처 엔지니어링 — 은 모두 "**역사의 데이터는 반복된다**"는 검증되지 않은 전제 위에 서 있다. 반복되지 않는다. 발견된 패턴은 발견되었기 때문에 사라진다(루카스 비판).
+
+정말로 반복되는 것은 데이터가 아니라 **데이터를 생성하는 세계** — 권한을 가진 기관, 규칙을 가진 규제자, 신호가 공개되어도 변하지 않는 인센티브 구조다. 그래서 이 프로젝트는 데이터의 잔여물이 아니라 생성기 자체를 모델링한다:
+
+> **우리는 데이터에서 구조를 발견한 것이 아니다. 데이터를 만들어내는 구조를 모델링했다.**
+
+두 가지 결론: (1) 모든 출력은 이름을 가진 주체·제약·균형 조건으로 귀속 가능 — 설명가능성은 부가 기능이 아니라 추론의 방향 그 자체. (2) 균형인 예측은 실행되어도 증발하지 않는다 — 실행되는 것이야말로 균형이 스스로를 실현하는 방식이다.
+
+## 하나의 엔진, 두 개의 프로덕트
+
+### To-C · 노이즈 제거 균형 가격 (중장기 보유용)
+
+정리 1(쌍대 크라메르-라오 한계): 행동 노이즈 ν_η는 누구도 거래로 이길 수 없다. 장기 투자자에게 정말 필요한 것은 노이즈 아래의 **균형 트랙 P^eq**과 괴리도 **D_t = P_t/P^eq_t − 1** — "당신이 사는 것은 가치인가, 과밀인가?"
+
+[](../../demo/denoised_price_2026.py)
+
+위 그림은 **합성 데이터** 컨셉 데모입니다 (2026년 7월 메모리 반도체 섹터 폭락 — SOX 지수 월간 −19%, 2008년 이후 최악 — 을 시나리오 원형으로 사용). 시뮬레이션에서 신호는 폭락 가속 **11거래일(약 2주) 전**에 점등. 실데이터 버전은 실험 E7로 명세화 완료.
+
+> *연구 신호이며 투자 자문이 아닙니다.*
+
+### To-B · 기관용 구조적 리스크
+
+Λₜ 레짐 모니터링(2008/2020급 신호), 포지션별 과밀도 분해, 이벤트 연산자 시나리오 분석. Airflow + Alpaca 파이프라인 스캐폴딩 완료.
+
+## 로드맵 — 1단계와 2단계
+
+**1단계 (현재)**: 게임이론 코어. HJB–FPK 연립계, 22개 이벤트 연산자, NeurIPS 워크숍용 실험 E1–E7. 데이터가 희소한 시대의 이론적으로 올바른 선택 — 정리 1은 잔차의 일부가 메커니즘으로*만* 설명될 수 있음을 증명한다.
+
+**2단계: [신경망 게임 구조(NNGS)](../PHASE2_NEURAL_GAME.md)** — **각 뉴런 자체가 작은 신경망**(기관·개인투자자 코호트당 하나)이며, 규제 현실이 아키텍처로 부과되고, 순전파 자체가 게임의 플레이다. 훈련 난도는 대략 제곱으로 증가하므로 H200급 컴퓨팅을 기다린다 — 설계 문서는 지금 공개한다.
+
+## 퀵 스타트
+
+```bash
+git clone https://github.com/hongjin-he/MicroWorld
+cd MicroWorld
+pip install -r requirements.txt
+
+python demo/global_demo.py # 애니메이션 월드 모델 데모 (약 2분, CPU)
+python demo/hindcast_2008.py # 2008 힌드캐스트
+python demo/denoised_price_2026.py # 노이즈 제거 가격 데모 (합성 데이터)
+python -m pytest tests/ -v # 테스트 스위트 (50개)
+```
+
+## 링크
+
+- **전체 README**: [English](../../README.md) · [中文](../../README_CN.md)
+- **수학 논문** (전체 증명 · 25쪽): [companion paper](https://github.com/hongjin-he/mathmatical-framework-for-world-models-in-quant-finance)
+- **개발 여정**: [docs/JOURNEY.md](../JOURNEY.md)
+- **파트너십·연락**: [LinkedIn](https://www.linkedin.com/in/hongjinhe-hkust-edu) · [X](https://x.com/Mr_Abstractor) · [GitHub issues](https://github.com/hongjin-he/MicroWorld/issues)
+
+> *컴플라이언스 고지: 이 저장소는 연구 소프트웨어입니다. 어떤 내용도 투자 자문이 아닙니다.*
+
+
+
+**Alpha Flow Research · HKUST · Stanford IHP · 2026년 7월**
+
+*이것은 퀀트 도구가 아니다. 금융 시장을 이해하는 새로운 패러다임이다.*
+
+
diff --git a/figures/denoised_price_demo.png b/figures/denoised_price_demo.png
new file mode 100644
index 0000000..d7afeac
Binary files /dev/null and b/figures/denoised_price_demo.png differ