Mario Marcolongo

AI Evaluation & Model Behavior Specialist

Model behavior testing · adversarial QA · evidence-bound reporting · evaluation operations

Italy · Italian/EU citizen · Open to worldwide relocation · International B2B contractingmariomarcolongo.com [email protected]LinkedIn GitHub

AI evaluation and research-verification specialist with self-directed model-behavior testing across chat, image, agentic tool-use and indirect prompt-injection challenges. The Gray Swan Proving Ground profile displayed rank #74 (top 6%) and 113 total breaks on 29 July 2026. Supporting work demonstrates multilingual content quality, consumer-genomics privacy research, corporate-source reconciliation, archival source recovery, claim-to-source auditing and evidence-bound reporting across sensitive records.

Selected evidence

113Platform-displayed Proving Ground breaks#74 · top 6% · 29 July 2026

255Arena submissions shown28 global unique breaks · 1,120 points

80Published content contributions55 YouTube videos · 4 articles · 21 short-form pieces

4,317Auditable Wikimedia contributionsEight years of inspectable claim and source work

Relevant experience

Model-Behavior Evaluator

Independent practice · Gray Swan Proving Ground

Evaluation record ↗Dated evidence ↗Public profile ↗

Jul 2026 — Present

Conduct self-directed testing of LLM instruction handling, policy boundaries and edge cases across chat, image, agentic tool-use and indirect prompt-injection settings.
Reached #74 on the Proving Ground (top 6%) with 113 platform-displayed total breaks in the 29 July 2026 snapshot.
Document the visible 112/113 discrepancy and separate platform-reported outcomes from independent verification, model-wide conclusions or security certification.

Scientific Content Quality & Operations Contractor

Entropy for Life · Independent contractor

Official Entropy for Life work record ↗

Jun 2023 — Present

Delivered 80 documented published content contributions: 55 YouTube videos · 4 articles · 21 short-form pieces.
Conduct recurring primary-literature review, scientific fact-checking and English-to-Italian localization; adapt evidence into Italian scripts, visualizations and short-form content while preserving terminology, meaning and source context.
Designed and built entropyforlife.it in WordPress; formally acknowledged in Giacomo Moro Mauretto's Mondadori book Italiani veri for scientific-literature research and error detection.

Additional relevant experience

Founder & Research-Workflow Owner

Yourself to Science™

Website ↗Repository ↗

Aug 2024 — Present

Founded and operate an open-source directory indexing more than 55 clinical studies, biobanks, donation programs, registries and other research initiatives.
Defined inclusion criteria, verification workflows, metadata structure, provenance requirements and licensing boundaries.
Define requirements, inspect code structure and behavior, test implementations and guide AI-assisted technical iteration through deployment and maintenance.

Additional evidence

Research-integrity product operations

Created and operate Notandia (formerly MDPI Filter), a browser and Zotero research-integrity tool. It helps researchers identify articles from publishers whose editorial and peer-review practices have attracted scrutiny—including MDPI and Frontiers—and checks Crossref/Retraction Watch records for formal notices such as retractions, corrections and expressions of concern. I define the evidence rules, privacy safeguards, ambiguity handling, false-positive boundaries and release tests.

Structured evaluation practice

Repeated adversarial testing across instruction hierarchy, multimodal inputs, tool-use behavior and untrusted external context.

Investigation and evidence-bound judgment

Attributed public cases demonstrate source synthesis, privacy-policy analysis, corporate-source reconciliation, archive recovery, legal-stage chronology, source-quality auditing, scientific taxonomy design and explicit separation of evidence from inference.

Capabilities

AI safety testingExploratory adversarial testing, prompt and jailbreak analysis, multi-turn behavior, multimodal inputs, agentic tool-use and indirect prompt injection

Evaluation operationsTest planning, evidence capture, reproducibility notes, taxonomy thinking, issue classification, severity-oriented reporting and mitigation-retesting concepts

OSINT and research verificationPublic-source and bibliographic research, claim decomposition, web-archive recovery, source-provenance analysis, corporate and legal record reconciliation, cross-source corroboration and evidence-bound reporting

Technical operationsCodebase reading and behavior inspection, requirements definition, functional testing, AI-assisted implementation workflows, Git/GitHub, JSON, REST APIs, WordPress, Cloudflare Pages and AWS Lambda deployment

Multilingual quality and communicationItalian-native and English-C1 content review, English-to-Italian scientific localization, terminology consistency, source-faithful adaptation and clear evidence limitations for technical and non-technical audiences

Best-fit role families

AI evaluation and safeguards operations
AI content red teaming and adversarial QA
Model behavior, trust & safety and policy testing
Human-data quality, grading and evaluation operations

Credentials & language

GALENOS Crowd Evidence Synthesis Training — Cochrane Crowd & GALENOS, 2026
Career Essentials in Generative AI — Microsoft & LinkedIn, 2024
EF SET English Certificate — 68/100, C1 overall, 2024
Italian — native. English — C1 overall (EF SET 68/100); advanced technical reading and professional writing, with practical English-to-Italian scientific localization and cross-language content-quality work.