Skip to content

v1.5.33

Release date: 2026-08-09

Summary

This release is a scientific-consistency and evidence-provenance pass across the IINTS-AF simulation, forecasting, calibration, MDMP, reporting, and documentation layers. It also ships the Rust/Tauri Research Workbench beta 0.2.6 and the maintained EUCYS technical dossier sources.

The release does not claim clinical validation. Its purpose is to make model assumptions, units, boundaries, uncertainty, and provenance easier to inspect and harder to misuse.

Patient Models and Units

  • Glucagon is converted explicitly between delivered milligrams and model plasma picograms.
  • Insulin delivery, insulin action, glucose mass, glucose concentration, meal appearance, sensor lag, and pump state use explicit and tested unit boundaries.
  • Bergman remote insulin action is driven by insulin relative to its reference state, preserving both increased and reduced action instead of forcing an always-positive effect.
  • Hovorka glucose compartments are initialized consistently with the configured basal state and retain mass-balance checks.
  • Invalid model domains now fail visibly rather than being silently clipped into plausible-looking values.
  • Snapshots remain backward compatible while recording the added research-state information.

HAAF, counter-regulation, renal clearance, exercise, circadian variation, tissue sensitivity, and stem-cell engraftment remain explicit research abstractions. Their code and parameter ranges are testable, but they are not substitutes for clinical or prospective validation.

Devices and Safety

  • Sensor lag and stochastic error are separated from blood-glucose state.
  • Pump delivery validation, reset behavior, occlusion handling, and state reporting are checked at the device boundary.
  • Meal coverage can use declared carbohydrates without allowing insulin-on-board logic to bypass low-glucose, predicted-low, or other deterministic safety guards.
  • Numerical or AI failures remain outside the deterministic safety authority.

Forecasting and Calibration

  • Physics-informed losses now use named raw physiological features rather than normalized tensor positions.
  • Unsupported universal rules such as "high IOB must always force a falling prediction" were removed; meal, insulin, timing, and uncertainty are evaluated as contextual signals.
  • Repeated stochastic inference can report epistemic uncertainty rather than presenting one neural output as certainty.
  • Evaluation emphasizes subject-level separation and reports fallback splitting when the dataset cannot support it.
  • Calibration adds held-out evaluation, parameter priors, physiological bounds, provenance, identifiability diagnostics, and explicit warnings for weakly constrained parameters.

These changes improve research discipline; they do not make a predictor suitable for dosing or real-time care.

MDMP and Provenance

  • Dataset population truncation is represented explicitly instead of disappearing from reports.
  • Contract and quality grades describe data checks, not clinical certification.
  • Evidence packages carry clearer source, configuration, environment, and limitation metadata.
  • Optional Ed25519 signatures support external integrity verification while preserving the existing hash-based manifest checks.

Formula Registry v5

The registry now distinguishes:

  • equations implemented from established model families;
  • IINTS adaptations with explicit implementation choices;
  • heuristic research mechanisms that require calibration or validation.

Formula identifiers, units, assumptions, code references, tests, and documentation are synchronized through the registry export. AI may explain those deterministic outputs but does not calculate or replace the registered equations.

EUCYS Dossier

Maintained source pages under docs/eucys provide:

  • Mermaid architecture and workflow diagrams;
  • the mathematical model and formula boundaries;
  • AI, data, MDMP, and reproducibility explanations;
  • benchmark figures with CSV source tables;
  • a live demonstration runbook, jury quick reference, and claim-to-evidence map;
  • builders for offline HTML, PDF, and review bundles.

Generated bundles are deliberately excluded from source control so every dossier can be rebuilt from versioned sources.

Rust/Tauri Research Workbench 0.2.6

The desktop beta was rebuilt and checked against SDK 1.5.33. It retains native path selectors, results inspection, local Ollama review, MDMP workflows, structural-biology tools, external-model labs, version diagnostics, and the private Python-engine installer.

Optional tools such as Ollama, COPASI, and OpenCOR remain external and are never installed or run silently. The native shell validates commands and paths before invoking the Python SDK bridge.

Verification

  • full Python suite: 878 passed, 2 skipped
  • focused scientific suite: 80 passed
  • desktop Python suite: 60 passed, 3 skipped
  • Rust/Tauri suite: 6 passed
  • mypy: no issues in 212 source files
  • flake8, governance checks, strict MkDocs, frontend static checks, Rust formatting, Clippy, release-mode build, and native --smoke passed

Safety Boundary

The release requires cryptography>=50,<51, which includes the upstream fix for PYSEC-2026-3552 affecting the previously permitted 49.x line.

IINTS-AF is pre-clinical research and educational software. It is not a medical device and must not be used for diagnosis, insulin dosing, glucagon dosing, treatment decisions, or real-time patient care.