Externalizing Epistemic Governance from Stateless Clinical LLMs
Philosophy & Technology, 2026. Sole author.
A constrained architecture for clinical decision support, with inspectable evidence, uncertainty and clinician overrides.
Stefano
NatangeloMD
Medical oncology.
Clinical AI research.
Software that puts research to work.
I’m a medical doctor completing oncology training at the University of Milan. I study how clinical language models can be evaluated and governed, and build software with explicit checks and traceable evidence.
Open to clinical AI evaluation and digital oncology roles from December 2026.
ESMO Merit Award 2026
Awarded for The Preservation Gap, on governing clinical language models across updates. First and presenting author, abstract #460, ESMO AI & Digital Oncology Congress.
Conference contributionsSoftware & systems
Tools for evidence verification and clinical workflows, built from research and day-to-day practice.
Callimachus
Citation verification with the evidence available for review.
I took Callimachus from a research question about AI governance to a publicly released desktop application. It checks reference identity, citation form and whether a source supports a manuscript’s claim.
Language models make bounded judgements. Deterministic checks validate the quoted passages and retain the evidence behind each finding.
Python, SQLite, PySide6. Desktop and command-line workflows; interchangeable model backends.
AIFA Registry Workflow Tool
A Windows application for AIFA registry documentation and read-only reconciliation with MedWeb. I built it for oncology trainees to reduce repetitive registry entry and flag duplicate or ambiguous records for human review. It includes manual takeover, local run histories and recovery after interruption.
Python, Playwright, Tkinter and SQLite.
Research in clinical AI
How to evaluate language models, preserve validated properties across updates and make clinical AI decisions auditable.
My independent work develops conceptual and methodological approaches to these questions. With AI-ON-LAB, I also work on clinical NLP and LLM benchmarking in oncology.
Selected work
Zenodo
2026
Preprint
Continuity Is All You Need: Why Clinical AI Governance Requires Structural Persistence
Argues that validated clinical AI properties must persist across updates and interactions.
arXiv
2025
Preprint · in revision
The Narrative Continuity Test: A Conceptual Framework for Evaluating Identity Persistence in AI Systems
Proposes five dimensions for evaluating an AI system’s continuity across time.
Communications Medicine
2026
Published
Clinician expertise and prompt engineering enhance cancer information extraction in electronic health records by small language models
Benchmarks four small language models on Italian oncology records and tests prompting strategies.
Clinical grounding
I work in thoracic oncology at Fondazione IRCCS Istituto Nazionale dei Tumori, with experience in immunotherapy and Phase I–III trials. My work includes patient screening and follow-up, safety reporting and structured clinical data.
My specialist training at the University of Milan also includes rotations at IEO, Niguarda and ASST Fatebenefratelli Sacco. Completion is expected in November 2026.
At Niguarda, I rebuilt the ward’s Word-based handover workflow; it remained in use after my rotation.
Full clinical experience and education
Journal of Secondary Endpoints
I founded and edit this diamond open-access journal on research culture and scientific satire, and built its publishing infrastructure. Through JSENetwork, I also provide pro-bono OJS hosting for independent journals.