EUCYS Jury Dossier¶

Project: IINTS-AF SDK
Author: Runebob Baers
Purpose: A browsable scientific and technical guide for EUCYS review
Software snapshot: IINTS-AF SDK 1.5.34

Research boundary

IINTS-AF is an open-source research and educational simulator. It is not a medical device, does not provide treatment advice, and must not control a real insulin or glucagon delivery system.

The project in one sentence¶

IINTS-AF is a transparent research workbench for creating virtual diabetes scenarios, testing experimental algorithms behind deterministic safety checks, and turning every run into evidence that another person can inspect.

How to browse this dossier¶

Chapter	What it answers	Best reader
Project and research question	Why was IINTS-AF built and what is being tested?	Everyone
System architecture	How do scenarios, physiology, algorithms, safety and reports connect?	Software and engineering judges
Scientific model and formulas	Which equations are actually implemented, and what are their limits?	Mathematics, biology and medical judges
AI, data and evidence	What may AI do, how is data handled, and where do external biology tools fit?	AI and data judges
Results, validation and limitations	What has been measured, what remains uncertain, and what must not be claimed?	Scientific judges
Live demonstration runbook	What should be shown and said during the demonstration?	Presenter
Jury quick reference	What are the short answers to likely questions?	Presenter and judges
Evidence map	Where can each claim be checked in code, tests, outputs and literature?	Reviewers

The 60-second explanation¶

Type 1 diabetes technology must react to glucose, meals, insulin already active in the body, exercise, sensor delay and uncertainty. Testing a new algorithm near a real person is therefore not the right first step.

IINTS-AF creates a virtual patient and repeatable scenarios. An experimental algorithm proposes an action, but a separate deterministic supervisor can reduce, block or flag that action. The SDK records the raw trace, settings, safety events, metrics and reports. Optional local AI may explain those artifacts, but it cannot invent measurements, solve the physiological equations or bypass the safety layer.

The scientific goal is not to prove that a controller is clinically safe. The goal is to make assumptions, failures and trade-offs visible before any clinical claim is considered.

What is implemented¶

Layer	Current capability
Virtual patient	Fast transparent model plus Bergman-style and Hovorka-style research ODE models
Disturbances	Meals, insulin, exercise, stress, circadian effects, sensor artifacts and experimental glucagon paths
Observation	CGM-like blood-to-interstitial lag, bias, drift, seeded noise and dropout
Algorithms	Baselines and experimental control or forecasting workflows
Safety	Deterministic input validation, dose constraints, intervention reasons and replay checks
Data	Import, standardisation, MDMP-style contracts, realism review and provenance artifacts
Evidence	CSV, JSON manifests, validation output, clinical-style reports, AGP-style assets and posters
Interfaces	Python SDK, CLI, Rust/Tauri desktop workbench and bench-only hardware adapters

What is not established¶

Clinical validity for treatment decisions.
Equivalence to a certified virtual-patient simulator.
Safety for real insulin or glucagon delivery.
Exact patient-specific physiology from a genetic, protein or tissue database.
Generalisable AI performance without independent external validation.
Proof that a realistic-looking curve represents a real patient.

Current evidence snapshot¶

The maintained EUCYS benchmark artifact contains 3600 locked simulation runs: six patient profiles, four scenario families, three study arms, five algorithm paths and ten seeds. Its main scientific value is the auditable matrix and its failure evidence, not a claim that every glucose outcome is clinically acceptable.

The aggregate arm table reports:

Study arm	Runs	Mean TIR 70-180	Mean time below 70	Mean time above 180
Clean, certified input	1200	90.30%	6.45%	3.25%
Corrupted, uncertified input	1200	73.19%	3.69%	23.12%
Supervisor-off ablation	1200	80.78%	3.88%	15.35%

These values require cautious interpretation. In particular, the low-glucose burden and high intervention counts are reasons for further investigation, not evidence of clinical safety. See Results, validation and limitations.

EUCYS benchmark overview