P01 · RQ01 · Review Wedge

基于观测数据进行因果建模的现状和思考

这篇综述不从“给定一个数据表,估计一个效应”开始,而从一个更前置的问题开始: 因果问题如何先于观测数据存在,观测数据又如何约束、校准和检验机制假设? 当前稿件已进入 v0.7:完整 1-8 skeleton、第一轮 citation/prose pressure、 Section 6 worked boundary example,以及第一张 question-first evidence bridge。

review_wedge paper.md v0.7 Figure 1 landed second example parked

当前 claim

观测数据不是因果建模的本体起点,而是约束、校准和检验机制假设的证据。 如果真实世界的个体机制本来就是异质的,综述结构应从平均效应中心转向 heterogeneity-as-primitive 的对象边界。

下一里程碑

做 public-facing readability polish,把 Figure 1 投影到 website / slide / LaTeX 格式,并继续保持 baseline-first:只有 causal artifact 改变时才使用 HCGM / DiscoSCM 作为主 claim。

当前证据状态

  • Question taxonomy 已进入正文结构。
  • HTE / CATE / ITE comparison 已作为 RQ23 支撑表。
  • F01 / Oregon LDTE 是 Section 6 主 worked boundary example。
  • 第二 worked example 已决策为暂不进入主文。

公开边界

这个页面是 public-safe projection,不发布内部草稿全文、私密 Discord 协作记录、 reviewer strategy 或 local-only source paths。事实源仍在项目 workspace 的 P01 artifacts 和 portfolio state 中。

Figure 1 · Question-first evidence bridge

P01 的第一张图把读者从“观测数据”带回“因果问题”:数据进入以后只是 evidence bridge, 不是因果关系本身。HCGM / DiscoSCM 只有在 artifact 或 decision 改变时才进入主 claim。 这里的 HCGM 比较是 object-level:标准 SCM 共享模型定义层 assignment,固定 exogenous state 后可诱导 unit-specific slice;HCGM / Unit Mechanism Learning 把 unit-specific mechanism 本身设为 primary object,并用 shared unit-modulated family 使它可学习,而不是声称 SCM 无法表达异质响应。

Question-first evidence bridge: causal question, observed world, evidence bridge, baseline artifact, primitive heterogeneity pressure, and paper-package test.
Layer Reader question Evidence role Failure test
Causal question before data What are we trying to explain, change, diagnose, or transport? World knowledge and counterfactual imagination define the candidate question. If the question is unclear, method choice is premature.
Observed world What did the table actually record? Sampling, selection, measurement, assignment, policy history, and behavior project the generated world into data. If the projection is ignored, association is mistaken for causal evidence.
Baseline causal artifact Which existing artifact already answers the question? PO, SCM/DAG, IV/LATE/LDTE, CATE/HTE, transportability, CRL, or dynamics may already be sufficient. If the baseline gives the same decision, do not invoke HCGM / DiscoSCM as the main claim.
Primitive heterogeneity pressure Which object remains unrepresented? Unit type, mechanism regime, event noise, context, measurement, or target-world relation may remain active. The added object must change the query, assumptions, failure mode, repair, or decision.

近期顶会 causality pressure

这些不是泛泛引用,而是从 top-ai-causality intake module 回流到 P01 的 reviewer pressure。

Distributional effects

NeurIPS 2025 `Beyond the Average` 让 P01 必须尊重 LDTE / IV / LATE baseline; 它是 Section 6 的主 worked boundary example,而不是被 HCGM / DiscoSCM 直接替代。

Counterfactual graphical baseline

ICML 2025 `Counterfactual Graphical Models` 压住一个边界:如果现有 counterfactual graph machinery 已经能表达目标 query,P01 就不能把 DiscoSCM 写成笼统替代。

Representation / dynamics sanity

ICML 2025 CRL sanity check、NeurIPS 2025 CausalVerse 和 CausalDynamics 共同要求: representation、simulator ground truth、dynamics 和 lag structure 必须进入 evidence standard。

Mechanism language boundary

FANS 是最强的未来 mechanism-shift second-example candidate,但在 final paper body / stable metadata / changed artifact 捕获前,先保留为 pressure lead。

当前 artifact set

  • `paper.md` v0.7 manuscript skeleton
  • `source-matrix.md` recent top-conference pressure bucket
  • `section-6-pressure-test-v0.md` IV / LATE / LDTE baseline pressure
  • `second-example-decision-v0.md` second worked-example decision
  • `figure-1-question-first-evidence-bridge-v0.md` first figure/table source

下一步判断

如果目标是对外可读,继续 P01 readability polish 和 projection;如果目标是 novelty evidence,切去 P17 的 benchmark-style comparison。当前默认仍是先让 P01 成为可展示 review wedge。

返回论文入口