Public ADAI manuscripts and a bilingual research website by Hiago Kin Levi, Decripte.
Website: https://hiagokinlevi.github.io/ADAI-Autonomous-Defensive-Artificial-Intelligence/
Short introductions to both papers are available in 15 languages. The English and Portuguese PDFs remain the authoritative manuscripts.
| Paper | DOI | Manuscripts |
|---|---|---|
| ADAI: Autonomous Defensive AI | 10.5281/zenodo.23089594 | English, Portuguese |
| DAI: A Domain-Specific LLM | 10.5281/zenodo.23089652 | English, Portuguese |
DAI is the proprietary domain-specific LLM within a governed defensive architecture. ADAI names the proposed technical category and formal framework. The two are not interchangeable. Original technical bodies are preserved in the mirrored PDFs.
Read the English guide or Portuguese guide for the connection to autonomous defensive AI, cybersecurity LLMs, information security, adversarial inputs and governed incident response. These explain the published papers and their limits; they do not introduce new experimental claims.
- English and Portuguese manuscript landing pages and searchable HTML reading views.
- The twelve mechanisms: AC, MEA, ACO, EBT, PPB, DSC, SAV, SSC, PPR, ELR, RBT and SRL.
- Model construction, architecture, reported observations and explicit limitations.
- DOI links, academic citation metadata, JSON-LD, canonical and language-alternate links.
- Related research with contextual links, without claims of endorsement or integration.
- GitHub preferred citation (CITATION.cff).
- BibTeX and RIS citations, Atom feed, XML sitemap and SHA-256 file manifest.
- Public records on Zenodo, OSF, OpenAlex, DataCite, OpenAIRE and Academia.edu, plus the educational-resource listing.
These are version 1.0 preprints prepared on 1 October 2026, not represented as peer-reviewed. Formal results depend on stated assumptions; historical observations are not universal performance guarantees. No third-party endorsement is asserted.
Static HTML/CSS with a small optional citation-copy script. No external font, tracking pixel, analytics, cookie, paid dependency or server-side application is needed. Publish the root of main with GitHub Pages. Serve locally with python3 -m http.server 8766.
Run python3 validate.py to verify local links, manuscript checksums, structured metadata and citation links. English and Portuguese content is rendered without JavaScript. PDF typesetting is authoritative; extracted text may not preserve mathematical notation, tables or diagrams.
All rights reserved. Public availability is not an open-source license and grants no rights to proprietary implementation or model weights. This repository contains the public research manuscripts, citations, glossary and website. Changes to manuscripts must be versioned consistently with their DOI records. Edit public HTML and metadata together and validate before publishing.
Contact: hiago@decripte.us