v1.5.5¶
Release date: 2026-05-17
v1.5.5 is the realism, EUCYS, and validation release.
This release collects the full SDK work completed after v1.5.4: more physiologically credible simulated glucose, a finished Tidepool import lane, stricter predictor validation, stronger Jetson endurance studies, a clearer first-run path, a rebuilt documentation experience, and a more polished EUCYS / live-demo package.
Highlights¶
1. Simulator realism is now grounded in empirical glucose behavior¶
The SDK now has a much stronger physiology layer for synthetic and simulated traces:
- physiological realism checks for implausibly flat traces
- causal meal-to-insulin realism checks
- empirical reference profiles and residual profiles derived from real CGM datasets
- new realistic day presets and calibration tooling
- dashboards and plots that compare simulated traces against real-data envelopes
The result is that simulated outputs are much less “toy-like” and much more useful for research review, demos, and model training.
2. Tidepool is now a real import path¶
iints import-tidepool is no longer a placeholder. It now:
- authenticates with a Tidepool session token supplied safely via environment variable or file
- resolves the current user automatically when
--user-idis omitted - imports CGM, bolus, wizard, and food events
- converts Tidepool glucose values into the SDK standard frame
- exports both
scenario.jsonandcgm_standard.csv
This makes the real-data toolchain much more complete alongside CareLink and Nightscout.
3. Predictor evaluation is much stricter¶
research/evaluate_predictor.py now supports:
- multiple external held-out datasets with
--external-data - feature-drift checks with
--reference-data - subgroup reports with repeated
--subgroup-column - hypo-detection sensitivity, specificity, false-positive rate, and missed-hypo rate
- MC-dropout uncertainty reliability bins and plot export
That means the SDK can now answer more serious questions than “does the LSTM run?”:
- does it generalize outside the training lane?
- does uncertainty mean anything?
- does performance degrade for specific cohorts?
- does it miss low-glucose events?
- has the evaluation distribution drifted from training?
4. Jetson endurance mode is safer for real long runs¶
Jetson studies now support:
- configurable simulated-time checkpoints
- partial
steps.csvpersistence during the run - periodic hardware telemetry in
raw/hardware_metrics.csv - richer status output with checkpoint minute, resume count, wall-clock elapsed time, and ETA
- more robust resume behavior after interruption
During release hardening, the resume path also uncovered and fixed a real sensor-state restore edge case caused by JSON tuple/list conversion.
5. One canonical onboarding command¶
New users now have one obvious route:
iints onboard
iints onboard --run-safe-steps
The command walks through:
- environment validation
- demo execution
- demo-data import
- realism checking
- study protocol generation
- study execution
That removes a lot of ambiguity from the first hour with the SDK.
6. EUCYS and live-demo packaging got much stronger¶
This release also includes:
- EUCYS PDF brief bundles
- clearer technical, physiological, ethical, and jury-facing materials
- a more usable live-demo CLI flow for calls and presentations
- easier edge onboarding for Raspberry Pi and UNO Q
- refreshed final evidence packaging
The public presentation layer is now much closer to the quality of the technical core underneath it.
7. Documentation was rebuilt instead of patched¶
The documentation site has been reorganized into a cleaner role-based structure with:
- a clearer homepage
- a developer portal
- generated API reference coverage
- better footer routing into technical docs
- cleaner onboarding, hardware, troubleshooting, and workflow paths
This is a real documentation rebuild, not a cosmetic sweep.
8. Smaller hardening work also landed¶
The release includes:
- local plugin foundations and versioned outputs
- runtime SQLite cleanup fixes
- governance and repository-root CI hardening
- dependency updates
- repeated notebook refreshes and health-badge maintenance
Validation¶
Before release, the updated SDK passed:
python3 -m pytest tests/ -q396 passed, 4 skippedflake8 .mypy src/iints/python3 tools/docs/generate_api_reference.py --checkPYTHONPATH=src python3 tools/ci/check_mdmp_sync.py
The full release check also rebuilds the docs, package artifacts, and technical manual.
Upgrade¶
Full workstation:
python -m pip install -U "iints-sdk-python35[full,mdmp]==1.5.5"
Edge / Raspberry Pi:
python -m pip install -U "iints-sdk-python35[edge,mdmp]==1.5.5"
Practical takeaway¶
v1.5.5 is the release where IINTS-AF becomes noticeably more convincing as a real research SDK:
- the data lane is broader
- the simulated physiology is more believable
- the AI evaluation story is stricter
- the long-run hardware evidence is safer
- the public documentation and EUCYS materials finally tell one coherent story