Cite the prior art the abstract was claiming past, and stop claiming generalization

[logannye]

Two paper-side defects, both of the kind a referee finds in the first minute.

1. The paper claimed to be something the repo says it is not

"A strict generalization of Pearl's structural causal models" opened the abstract and the conclusion — while §1.2, two paragraphs later, said the contrary ("the contribution is not new expressive power") and README.md said "It is not more expressive." The sentence a referee reads first contradicted the repository's own position, and nothing could catch it.

Both now say reformulation, and say plainly why: Pearl with sufficient latents is universal, the two operations are σ-interventions of a restricted form, and under C1–C4 the identifiability results are corollaries of Tian-Pearl c-component factorization (Prop T4.0) rather than independent theory.

2. The nearest prior art was uncited

work	why it matters
Shpitser & Tchetgen Tchetgen (2016), Ann. Statist. 44(6)	Places node interventions inside edge interventions; gives the edge g-formula and the recanting witness. Mechanism-level identification theory is a decade old
Ma et al. (2022), KDD	Causal inference on hypergraphs — so "these systems are descriptive" was false as written, and a copy sat untracked in the repository root while the abstract claimed the field was empty
Yao & Evans (2022)	Edge interventions in linear SEMs, necessary & sufficient conditions
Correa & Bareinboim (2020)	§2.1.1 already argued against σ-calculus without citing it

§2.1.1 now adds the sentence that should always have been there — we do not claim priority on intervening at the level of a mechanism — and states what actually is ours: the typed hypergraph carrier (their edge is dyadic, so a mechanism with several jointly-produced outputs is a set of edges rather than one object) plus a compiler emitting estimand, assumptions and refusal as machine-checkable artifacts.

§2.4 now distinguishes Ma et al. rather than ignoring them: their hyperedges carry interference between units, ours are the mechanisms within one, and no intervention in their setting addresses a hyperedge as the object being changed. The surviving claim is the narrower one, about hypergraph SCMs.

⚠️ Found by auditing every in-text citation against the list

Two dangling citations: Eberhardt & Scheines (2007) and Tian & Pearl (2001), both cited in §2.1.1 with no reference entry. Added. The list is now 26 entries and alphabetical — checked mechanically, because two of my own insertions broke the ordering and one spliced Yao & Evans into the middle of the Geiger-Verma-Pearl entry by matching a surname mid-line.

§2.4's "causal hypergraph transformer" also gains a caveat: whether hypergraph structure is load-bearing for a trained model is a different question from whether it is load-bearing for identification, and a negative answer to the first leaves this paper's results untouched.

Verification

tests/test_whitepaper_claims.py gates all of it, for the reason test_readme_smoke.py exists.

mutation	result
reinstate "strict generalization" in the abstract	2 tests red
drop the Shpitser & Tchetgen Tchetgen reference	3 tests red

The citation audit and the alphabetical check both found real defects while being written, which is the argument for keeping them.

.pdf is gitignored — reference copies are read locally, never redistributed from here. README.md's C1 claim was already correct (lines 503–504) and needed no change.

339 passed, 1 xfailed. ruff check . clean.

🤖 Generated with Claude Code