Journals closest in thematic statistical profile typically share no topics: information-preservation requirements for representing journal thematic identity

Journals are routinely represented thematically by a single object — a subject category, a diversity index, an embedding vector. The prior question those representations presuppose is rarely asked: what information must a representation preserve in order to be faithful to what a journal actually publishes? This preprint answers that question empirically, by using reduction experiments on article-level thematic observations to determine what information cannot be discarded. A domain of thematic properties is fixed before any reduction is attempted. Each candidate representation is then treated as a hypothesis about what may safely be discarded, and is retained only where its removal produces a categorical violation or a measured, uncompensated loss on pre-declared instruments. The design was run in two stages: an exploratory stage that derived six information-preservation requirements, and a pre-registered confirmatory stage that tested a lattice of candidates against requirements fixed and sealed before the confirmatory data existed. The reductions fail in specific and measured ways. A profile carrying distributional form and article-level organization, but no labelled topic identity, places beside each journal–year cell a nearest neighbour that shares a median thematic mass of zero — exactly zero for 58,670 of 108,000 queries — with neighbourhood content overlap 0.214 against 0.943. Article-level organization is not recoverable from aggregate composition (RVR 0.738 against a 1.2 × 10-4 ceiling). Content- and organization-induced neighbourhoods share no members for the median query in 197 of 216 slices. Taken together — and not individually — these results constrain the admissible representation, and the paper formalizes a conforming representation, I_j, that preserves the constrained structures jointly, without fusion. Scope. The relational findings, and the irreducibility claims resting on them, are declared valid for the instantiation examined, pending a venue-ablated re-instantiation of the observation layer or a cross-infrastructure reproduction. The paper claims no statistical independence between the two relational views: no test against a null of independence was performed. It falsifies an examined lattice of candidate representations, not the space of all conceivable ones. Data, code and the full reproducibility record are deposited separately at doi:10.5281/zenodo.22014623. Every reported quantity is traceable by SHA-256 to a sealed execution artefact. This is the author's manuscript, prepared for submission to Quantitative Science Studies. It has not been peer reviewed.