Jury Quick Reference¶
One-sentence pitch¶
IINTS-AF is an open-source research workbench that makes diabetes-algorithm experiments reproducible, safety-aware and inspectable before any real-world claim is made.
Thirty-second answer¶
The SDK creates a virtual patient, adds meals, insulin, exercise and sensor effects, and lets an algorithm propose simulated actions. A separate deterministic supervisor checks those actions. Every run writes the raw trace, settings, safety events and reports. Optional AI explains the result but has no numerical or dosing authority.
Five facts to remember¶
| Fact | Short answer |
|---|---|
| What is it? | A pre-clinical simulation and research SDK |
| What is the key design? | Proposal and safety permission are separate |
| How many registered formulas? | 15 deterministic formulas in registry v5 |
| What is the EUCYS experiment? | Normal, stress and risk scenarios plus a larger locked benchmark |
| Is it medical software? | No; research and education only |
Likely questions¶
What did you personally build?¶
An integrated SDK and desktop workbench covering virtual-patient simulation, algorithm interfaces, deterministic safety checks, data validation, AI research workflows, reports, evidence bundles and bench-hardware adapters.
What is scientifically new?¶
The contribution is the combination of transparent physiology, explicit candidate-versus-safety authority, data-quality gates and reproducible evidence packaging in one open research workflow.
Why not use one existing simulator?¶
Existing simulators remain important references. IINTS-AF focuses on an open, inspectable workflow around algorithms, safety, data, reporting, AI and hardware. It must still be compared with independent reference simulators.
Does the AI calculate the formulas?¶
No. Python code evaluates the registered equations. The language model only receives fixed formulas and recorded metrics as explanation context.
Can the AI deliver insulin?¶
No. An experimental controller can propose a simulated action, but deterministic safety logic remains the final simulated authority. The SDK must not control a real pump.
How do you know the physiology is realistic?¶
The model is grounded in Bergman-, Hovorka- and meal-model literature and has plausibility checks. It is still a research approximation. Stronger held-out real-data calibration and independent validation are explicit next steps.
What is the strongest result?¶
The platform executed a locked 3600-run matrix and exposed measurable differences between clean and corrupted inputs. Equally important, it exposed low-glucose and intervention burdens that require further research.
Why report bad outcomes?¶
Because a scientific test platform should reveal failures rather than optimise the presentation. A negative or concerning result helps identify the next experiment.
What does MDMP add?¶
It checks an explicit data contract, provenance and quality before data is used as evidence. Its certificate documents the check; it is not a clinical approval.
What do AlphaFold and ClinVar add?¶
They add structural and variant-classification context. They do not automatically determine patient insulin sensitivity or a quantitative disease effect.
How is the work reproducible?¶
Runs record patient, scenario, algorithm, duration, step, seed, software version, raw trace, validation output and file manifests.
What would you do next?¶
Investigate low-glucose failure clusters, normalise safety interventions by episode and time, calibrate against held-out subjects, and compare with an independent reference simulator.
Four formulas to explain live¶
Meal delay:
Insulin delay:
Hovorka glucose mass:
Sensor lag:
Demo command¶
iints demo eucys \
--output-dir results/eucys_live \
--skip-ai \
--evidence
Safe closing line¶
IINTS-AF does not prove that an algorithm is safe for people. It makes a pre-clinical experiment easier to reproduce, inspect and challenge.