Superquadratic histories and certified conditional training in hard leaky integrate-and-fire networks: source data and code

Abstract

This record contains source data and original software accompanying “Superquadratic histories and certified conditional training in hard leaky integrate-and-fire networks”. It retains construction records, feasible-history candidates and exact prices, frozen model and prediction arrays, numerical audit logs, figure source data, reproducibility protocols, and unsuccessful or unfavorable results. The work studies hard delayed-reset LIF dynamics, temporal-history complexity and conditional certificates on fixed robust domains; it does not establish predictive superiority or efficient unrestricted high-dimensional training. Original software is licensed under MIT. Generated results, figures and original reproducibility documentation are licensed under Creative Commons Attribution 4.0 International. Third-party rights remain unchanged. Raw GunPoint TRAIN/TEST files and full derived input encodings are excluded; pinned source URLs, frozen hashes and restoration code are included. Manuscripts, independent proof drafts, their page renderings, extracted transcripts, and embedded narrative sources are excluded from this release. Full proofs accompany the journal manuscript as Supplementary Material. Some preserved historical QA and figure wrappers depend on omitted local proof texts for provenance freezing. Their standalone rerun restrictions and supported integrity/byte-restoration entry points are documented in metadata/QA_REPRODUCTION_LIMITATIONS.md and metadata/static_dependency_map.json. The ZIP preserves its predeposit build metadata. Statements inside it describing a local candidate with no DOI refer to the state at archive creation; this Zenodo record provides the current publication status and persistent identifier. OpenAI Codex/ChatGPT assisted mathematical exploration, literature work, coding, execution and checking, manuscript drafting and editing, and programmatic figure preparation. The author directed the work and accepts responsibility for the content. Tool-assisted checks do not constitute independent human verification.