USL10 Paper Portfolio
USL-03 Counterfactual · Treatment Effect registered slot · current semantic abstract local only public contract: drafting Primary deep-work lane

Same-Unit Counterfactuals: Identification Beyond Fixed Unit Beliefs

When does factual evidence support a counterfactual query about the same actual unit?

venue AAAI-27 Main Track topic RU: Causality state abstract registered / writing owner gong · CausaClaw × DiscoSeed OpenReview forum 待回填

Semantic-revision boundary. The owner-side OpenReview Activity audit records the abstract slot, but the current local working abstract contains a later semantic revision (shared_u_identification_revision_not_verified_external). The direct OpenReview forum URL is also still pending writeback.

基本研究问题

这篇论文独立可证伪的问题是什么；它与共享 ontology 中其他九篇不重叠的部分。

When does factual evidence support a counterfactual query about the same actual unit?

核心 claim
World-side shared-$U$ semantics, not equality of learner beliefs, keeps a counterfactual about the same already-selected individual; factual evidence first defines $P(U\in du\mid\mathcal O^F)$ and the learner forms $Q_\phi(du\mid\mathcal O^F)\approx P(U\in du\mid\mathcal O^F)$, while keeping the candidate intervention outside $\mathcal O^F$ only enforces a no-query-leakage learner protocol and does not identify a conditional counterfactual without the required estimand, support and structural or cross-world event-coupling assumptions.
如果成立会改变什么
Counterfactual systems must report whether they answer for the same unit, a same-mechanism donor or only a population analogue.
最强 reviewer 反对 / kill signal

反对：The result is a restatement of standard abduction-action-prediction and familiar non-identification.

Kill signal：kill if the paper cannot add a selection-specific theorem, estimand or diagnostic beyond existing SCM non-identification results.

基本思路

论文如何回答这个问题：形式化对象、论证 spine 与所需证据。

USL-01 type contract

The Unit Selection Variable $U\sim\Pi$ performs one world-side individual selection under the task-declared Population law $\Pi$, and the same random variable is shared across factual and alternative worlds; on $\{U=u\}$ both worlds concern the same realized individual. Factual evidence first conditions the Population world law $P(U\in du\mid\mathcal O^F)$; the answer-time learner forms $Q_\phi(du\mid\mathcal O^F)\approx P(U\in du\mid\mathcal O^F)$. The candidate intervention $a'$ is not part of $\mathcal O^F$, so this learner belief is reused rather than recomputed from the candidate query. This is a no-query-leakage computation rule, not same-individual linkage and not a causal identification assumption. A factual-outcome-conditioned counterfactual such as $P(Y_U(a')\in dy\mid\mathcal O^F)$ can require restrictions on the joint factual/counterfactual event law; a fixed-$u$ marginal intervention response is a different estimand and need not require the same coupling. If $u\neq v$ but $Z_u=Z_v$ or their response laws agree, they may be same-mechanism donors but remain different individuals.

论证 spine

  1. Why similar-unit prediction is not a same-unit counterfactual.
  2. One realized Unit selection shared across factual and alternative-action worlds.
  3. Factual evidence, non-evidential candidate intervention and no-query-leakage learner protocol.
  4. World-side same-individual linkage versus learner $Q_\phi$ accuracy.
  5. Marginal intervention response, factual-outcome-conditioned counterfactual and joint-effect estimands.
  6. Structural/event coupling, non-identification theorem and sharp partial-identification sets.
  7. Oracle simulation: same individual, same-mechanism donor and population analogue.
  8. Reporting standards and abstention rules for individual counterfactual systems.

所需证据

  • Observationally equivalent models with different individual counterfactuals.
  • At least one non-vacuous bound under a declared coupling class.
  • A diagnostic that catches re-selection of the unit after changing the query.
  • A two-model construction matching selector laws, shared-$U$ semantics and fixed-$u$ one-world action marginals while changing only the relevant cross-world event coupling.
  • Separate the causal identification theorem from calibration/estimation error in $Q_\phi$, and type same-mechanism donors through $Z_u$ or response-law equivalence rather than individual identity.
  • State and evaluate the learner approximation $Q_\phi(du\mid\mathcal O^F)\approx P(U\in du\mid\mathcal O^F)$ separately from the shared-$U$ world semantics and the causal identification assumptions.
Current working abstract（本轮语义修订尚未验证已同步外部记录）

We distinguish world-side same-individual semantics from the learner protocol used to answer a counterfactual query. Sharing one Unit Selection Variable $U\sim\Pi$ and its realized value $u$ across factual and alternative worlds keeps the query about the same realized individual under the task-declared Population law $\Pi$. Factual evidence first defines the world conditional $P(U\in du\mid\mathcal O^F)$, and the learner forms $Q_\phi(du\mid\mathcal O^F)\approx P(U\in du\mid\mathcal O^F)$. Reusing this belief while keeping the candidate intervention outside $\mathcal O^F$ merely prevents query leakage. Identification then depends on the declared estimand: factual-outcome-conditioned individual counterfactuals may require structural or cross-world event coupling, whereas a fixed-individual marginal intervention response is a different target.

当前进展

状态只记录可验证 delta：claim、SOTA opponent、theorem、experiment、manuscript 或 owner decision。

当前状态

AAAI-27 abstract registered; full paper drafting.

下一写作门

challenge the binary matched-model construction and sharp coupling bounds against prior art; add a genuinely selection-specific theorem/diagnostic or trigger the existing kill/merge signal. Do not relabel the baseline cross-world non-identification fact as new.

下一证据门

include an oracle simulation separating same unit from matched donor and a diagnostic that detects alternative-query-induced Unit reselection.

48h 最小实验

construct two models matching the factual observed law, task-declared Population law, world conditional over the Unit Selection Variable, shared-$U$ semantics and fixed-$u$ one-world action marginals, but differing in factual/counterfactual event coupling and therefore in a conditional same-individual counterfactual; derive sharp bounds under declared coupling classes.

Closest SOTA opponents

structural counterfactual abduction; potential-outcome individual effects; DeepSCM/CEVAE-style latent counterfactual models.

canonical 文件

research-questions/USL03-same-unit-counterfactuals/seed.md
papers/USL03-same-unit-counterfactuals/paper.md

进展日志

2026-07-29 · current local semantic abstract revision shared_u_identification_revision_not_verified_external has not been verified as synchronized to the external record; the registered slot remains tracked separately.

后续每次可验证 delta 追加在此；canonical 状态以 seed.md / paper.md / submission-ledger 为准，本页由 build_paper_pages.py 重新生成。