Answers are grounded in this paper and cite their sources.
Read the paper: Manuscript PDF. Explore the predictions: Complete prediction/error atlas. Preprint and reproducibility archive. This work has not undergone journal peer review. An incorrect final answer need not indicate an incorrect computation, and high cell agreement need not indicate useful prediction. This retrospective study examines ten hosted language-model endpoints on explicitly supplied elementary cellular-automaton Rules 30 and 110: 126 shared cases, three workflows, and 4,650 primary and supplementary response records collected on 23 September 2026. Interactive Python produces protocol-compliant exact answers in 94.3% of primary trials, compared with 3.5% for answer only and 2.7% for written working without execution. The analysis separates correct final content, response-format and protocol requirements, successful execution, and visible tool results. Twenty-two unsuccessful Python trials expose the correct target in tool output; 31 other unsuccessful trials execute successfully but expose no output. A conservative content diagnostic identifies 11 additional exact Python answers without changing the original protocol scores. Further analyses account for reasoning-setting score changes through gained and lost eligible answers, distinguish unchanged-background agreement from accuracy inside the evolving region, and report repeated-success distributions and observed inference costs. All models and workflows are shown together, with a complete 14-page prediction/error atlas. The comparisons use unequal workflow resources and a small, correlated task family. Requested API reasoning settings do not establish the absence of internal reasoning. Results describe the tested service endpoints and measurement interface, not proprietary neural mechanisms, equal-compute causal effects, or universal model rankings. All Qwen observations are API measurements; there was no local GPU experiment. Contents: named manuscript PDF; prediction atlas; self-contained LaTeX source; code, configurations, archived requests and responses, execution evidence, analysis outputs, offline reproduction instructions, and checksums. No new inference was performed for this release. The independent human-review annotation packet is provided but remains uncompleted. AI assistance: OpenAI Codex was used extensively under human direction for implementation, experiments, analysis, literature research, plots, and writing. The manuscript gives the full disclosure; AI systems are not authors. Rights: the paper and original documentation and figures are CC BY 4.0. Original benchmark software is MIT licensed. Third-party components and source/service outputs retain any applicable rights and terms; the archive includes a scoped rights statement.
No takes yet. Share an insight, caveat, or question.
Ed Herranz (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: