--- title: "AI Safety Formalization Atlas — Mario Marcolongo" description: "Five merged contributions to source auditing, citation provenance and repository tooling, and coauthorship of the AI Safety Formalization Atlas preprint." canonical: https://mariomarcolongo.com/work/atlas --- [Selected work](/work) - Open-source contributor - Preprint coauthor - 2026 # Making research sources traceable Five merged contributions to source auditing, citation provenance and repository tooling; coauthor of the Atlas preprint released on Zenodo on 1 October 2026. [Read the preprint · DOI](https://doi.org/10.5281/zenodo.23088952)[Source audit · PR #69](https://github.com/mbrcic/ai-safety-formalization-atlas/pull/69)[Source locators · PR #70](https://github.com/mbrcic/ai-safety-formalization-atlas/pull/70)[Accepted PR #52](https://github.com/mbrcic/ai-safety-formalization-atlas/pull/52)[Accepted PR #54](https://github.com/mbrcic/ai-safety-formalization-atlas/pull/54)[Source-backed agent-instruction guide](https://github.com/mbrcic/ai-safety-formalization-atlas/blob/66ceba19c1904f00b2a734486ee0de3f5fc5a470/docs/guide/agent-instruction-design.md)[Accepted PR #61](https://github.com/mbrcic/ai-safety-formalization-atlas/pull/61) [![GitHub pull request 70 showing Mario Marcolongo's accepted source-locator contribution and merged status](/media/work/atlas-pr-70.webp)](https://github.com/mbrcic/ai-safety-formalization-atlas/pull/70) Actual merged PR #70: source locators and publication provenance · captured 2026-09-30. ## A usable source trail Formal mathematics still depends on knowing which source, version and statement a result refers to. My recent Atlas work makes bibliographic comparisons inspectable and references easier to follow. The public contribution record now contains five merged PRs, each with its own scope and review trail. ## Source metadata and rights-evidence audit PR #69, merged on 22 September 2026, adds a reproducible audit that compares catalogue entries with Crossref, arXiv and public-page metadata. It saves dated evidence and generates a report of differences, missing information and retrieval failures for human review. Bibliographic comparison is separate from registry validation; licence observations remain distinct from a rights determination. The PR records a 109-work snapshot and full local validation. Those are dated audit and validation results, not counts of research I personally formalized. ## Missing locators and publication provenance PR #70, merged on 28 September, adds 15 missing source locators and enriches selected references. It preserves the cited technical report where a later journal version also exists, with supplementary publication information. My work included verifying locators and statement correspondence for selected sources. Automated retrieval failures remain visible; they are not treated as proof of an incorrect citation. ## Earlier repository contributions PR #52 diagnosed and repaired a Bash 3.2 validation failure on macOS and improved migration checks. PR #54 added portable instruction adapters. In PR #61, I researched official platform guidelines, instruction discovery and precedence, context limits and maintenance practices; wrote a 312-line cross-platform agent-instruction guide; and reorganized repository guidance into task-scoped policy files. These earlier contributions address tooling and repository instructions. Their public diffs describe the specific changes and validation limits. ## Public preprint and affiliation The AI Safety Formalization Atlas: a machine-checked memory for AI safety mathematics is publicly available as a Zenodo preprint, version 1, released on 1 October 2026. The manuscript is dated September 2026. Its title page lists me as a coauthor alongside Mario Brčić, Luka Hobor, Mihael Kovač and Adrian Satja Kurdija. Metropolia University of Applied Sciences is my listed affiliation, used with permission; I undertake Information Technology path studies through its Open UAS. The public record establishes preprint authorship, not peer-reviewed publication. ## Individual contribution and project scope The linked PRs establish my accepted work on source quality, provenance and tooling, with AI-assisted implementation. I do not attribute the Atlas’s library size, catalogue totals, formal proofs or mathematical findings to myself. The public manuscript establishes coauthorship; the accepted PRs document my individual contribution scope. Review the public diffs for the concrete scope of my work. ## Research profiles [ORCID](https://orcid.org/0000-0003-2846-7115) · [Google Scholar](https://scholar.google.com/citations?user=EnNykUIAAAAJ&hl=en) · [ResearchGate](https://www.researchgate.net/profile/Mario_Marcolongo) The Atlas preprint and software archives link to the original records on their project pages. ## Claim provenance Contributed merged PR #52: migration validation portability. Acceptance covers the submitted change; it does not establish authorship of the wider research or independent engineering proficiency. Evidence: external acceptance · Owner: upstream project · Observed 2026-09-12 · Event/start 2026-09-03. [Original evidence](https://github.com/mbrcic/ai-safety-formalization-atlas/pull/52) [In the evidence index →](/evidence#atlas-pr-52-accepted) Contributed merged PR #54: portable repository instruction adapters. Acceptance covers the submitted change; it does not establish authorship of the wider research or independent engineering proficiency. Evidence: external acceptance · Owner: upstream project · Observed 2026-09-12 · Event/start 2026-09-03. [Original evidence](https://github.com/mbrcic/ai-safety-formalization-atlas/pull/54) [In the evidence index →](/evidence#atlas-pr-54-accepted) Contributed merged PR #61: repository instruction organization. Acceptance covers the submitted change; it does not establish authorship of the wider research or independent engineering proficiency. Evidence: external acceptance · Owner: upstream project · Observed 2026-09-12 · Event/start 2026-09-06. [Original evidence](https://github.com/mbrcic/ai-safety-formalization-atlas/pull/61) [In the evidence index →](/evidence#atlas-pr-61-accepted) Contributed merged PR #69: a human-reviewable source metadata and rights-evidence audit. The audit compares metadata and preserves retrieval evidence for human review; it does not certify citations or reuse rights. Most added lines are generated reports and evidence, not handwritten tooling. Evidence: external acceptance · Owner: upstream project · Observed 2026-09-30 · Event/start 2026-09-22. [Original evidence](https://github.com/mbrcic/ai-safety-formalization-atlas/pull/69) [In the evidence index →](/evidence#atlas-pr-69-accepted) Contributed merged PR #70: 15 missing source locators and clarified publication provenance. Acceptance covers source traceability and bibliographic changes; no Lean theorem, proof or formalization-coverage change is claimed. Evidence: external acceptance · Owner: upstream project · Observed 2026-09-30 · Event/start 2026-09-28. [Original evidence](https://github.com/mbrcic/ai-safety-formalization-atlas/pull/70) [In the evidence index →](/evidence#atlas-pr-70-accepted) Coauthor of “The AI Safety Formalization Atlas: a machine-checked memory for AI safety mathematics,” publicly available as a Zenodo preprint since 1 October 2026. Public preprint, not an established peer-reviewed publication. Documented individual contributions concern source auditing, citation provenance and repository tooling; authorship does not imply responsibility for all proofs. Metropolia is the listed affiliation, with permission confirmed by the contributor; no employment or degree is implied. Evidence: public preprint · Owner: project authors · Observed 2026-10-14 · Event/start 2026-09. [Original evidence](https://doi.org/10.5281/zenodo.23088952) [In the evidence index →](/evidence#atlas-paper-preprint) Source review: 2026-10-14. This is a dated account; the linked sources may have changed.