Rayan Pal · Zenodo (CERN European Organization for Nuclear Research) 2026 · 2026
DOI: 10.5281/zenodo.22983233
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
This paper reports two preregistered experiments testing whether a natural-language prerequisite relation controls observable continuation in gpt-5.4-2026-03-05. Across 1,600 fresh single-turn requests, the experiments manipulate the presence of a binding construction and the linguistic object assigned prerequisite status. Experiment 1 crosses the binding rule with two Arabic inputs. When the rule names شَرْط as the prerequisite, the matching input produces visible continuation in 200/200 trials, including 185 exact cross-script outputs شָׁרְט. The mismatching input شَمْس produces 200/200 provider-successful, normal-stop responses containing zero visible UTF-8 output bytes, classified as V0. Removing the binding rule eliminates V0 in both input conditions: 0/200 each. The preregistered interaction contrast is 1.000. Experiment 2 retains the binding syntax and exchanges prerequisite status between the same two Arabic objects. Matching conditions produce visible continuation in 400/400 trials. Mismatching conditions produce V0 in 398/400 trials: 200/200 under the original assignment and 198/200 under the reversed assignment. Both nonblank exceptions remain in the evidence and analysis. The preregistered crossover contrast is 1.990. Both experiments satisfy their respective preregistered support criteria. All 1,600 primary requests succeed with the exact requested model identifier, without retries or exclusions. The public artifacts preserve frozen prompts, Unicode identities, request schedules, runtime parameters, raw provider responses, classification rules, statistical analyses, hashes, and deterministic offline verification. The manuscript includes exact request constructions, complete outcome tables, raw witnesses, and separate inferential analyses for the two experiments. Together, the experiments establish rule-dependent and role-reassigned linguistic prerequisite gating within the tested construction. The observed continuation boundary follows the prerequisite relation assigned in language rather than either Arabic object in isolation. The experiments do not identify an internal stopping mechanism or establish generalization across synonyms, translations, or arbitrary predicates. Paper, source, figures, supplement, and build instructions:https://github.com/theonlypal/gpt-5.4-prerequisite-role-reassignment-paper Experiment 1 evidence and analysis:https://github.com/theonlypal/gpt-5.4-shrt-prerequisite-gate-analysis Experiment 2 evidence and analysis:https://github.com/theonlypal/gpt-5.4-shrt-prerequisite-reassignment-analysis
No comments yet — start the discussion below.