Structure substitutes for scale in cross-document temporal reasoning

Abstract

Retrieval systems are judged on whether the right evidence reaches the model. This paper measures what happens when it does and the model still cannot answer, because the answer is a relation between passages rather than a statement inside any of them. Over a five-document narrative corpus (260,204 words; 13,950 passages), 38 questions ask whether one event precedes another when the two are narrated in different documents and share no character, place or causal link. No passage states either relation. Five models of one family (Qwen3, 0.6B–14B) answered each question twice: given the source passages as text, and given the identical facts as a structured chronology block built from an explicit state store. Given the passages, every model scored 0/38 and refused 92–100% of the time — correctly, since the ordering is genuinely absent from the text. Given the same facts as structure, an 8B model scored 28/38 (73.7%). A four-condition ablation separates information from form. At 14B the form is irrelevant: plain prose, sorted prose and a structured block all reach 73.7%. At 8B the form is worth +32 percentage points. An 8B model given structure matches a 14B model given prose. Structure and parameter count are, over this task, partially interchangeable. Permuting the supplied story positions collapses accuracy to 10.5% (8B) and 21.1% (14B), establishing that the models follow the supplied ordering rather than recalling the published narrative — a contamination control. One failure is universal: no model, at any size, under any representation, concluded simultaneity when two events shared a story position (0/10 in 11 of 12 cells). The protocol was pre-registered before any model was run. A prior headline result from this project (86%) was withdrawn after inspection and is reported here as withdrawn. Author. Panagiotis Gkilis, founder, BedVibe Studios (Oslo, Norway). Portfolio: tts.bedvibe.studio/portfolio. Engineering notes: ai.bedvibe.studio. ORCID: 0009-0007-3805-170X.