v1.5.7¶
Release date: 2026-05-22
v1.5.7 is the Jetson physiology and local-AI research release.
This patch turns long Jetson runs into a clearer research pipeline: acquire a real-duration bundle, audit physiological plausibility, export training datasets, train local controller/predictor models, and evaluate learned controllers on held-out simulation scenarios.
Highlights¶
1. Jetson physiology telemetry is stricter¶
Endurance runs now report whether the algorithm-facing CGM stream diverges from the simulated patient truth:
- physiology warning count in
status.json - input-validator fail-soft rows
- algorithm-blind hyperglycemia rows
- mean absolute truth/sensor gap
- a dedicated
Physiology Qualitysection infinal/ENDURANCE_REPORT.md - main figures that can show mechanistic truth, reported truth, and algorithm input separately
This makes it much harder for a run to look acceptable while the model, sensor, or validator is hiding physiologically important behavior.
2. Safer physiology defaults¶
The educational custom-patient defaults no longer use an overly aggressive glucose decay rate. Clinic-safe virtual-patient YAMLs were also relaxed so quickstart and endurance paths are less likely to create unrealistic pre-meal glucose crashes.
3. Long runs now export real local-AI training artifacts¶
Jetson endurance research bundles now preserve both the observed controller output and a conservative research teacher label:
observed_delivered_insulin_unitsreference_teacher_insulin_unitsteacher_insulin_unitscontroller_teacher_policy: conservative_reference_v1
The teacher is explicitly research-only and is designed to be safer than copying weak demo-algorithm output as the training target.
4. New iints research local-ai-lab¶
A new command combines one or more completed Jetson/simulator runs into a local AI workspace:
iints research local-ai-lab \
--run day1=results/jetson_research_day \
--run day2=results/jetson_research_day_2 \
--output-dir results/local_ai_lab
It writes:
datasets/predictor_training.csvdatasets/controller_teacher_dataset.csvdatasets/LOCAL_AI_DATASET_CARD.jsonmodels/linear_controller.json- optional
models/neural_controller.pt - optional predictor-training output
- optional held-out controller evaluation
LOCAL_AI_RESEARCH_SUMMARY.jsonLOCAL_AI_RESEARCH_REPORT.md
This is the recommended path when the goal is to use long Jetson runs as training material for local AI research.
5. Packaging polish¶
The Typer dependency no longer requests the obsolete typer[all] extra, avoiding noisy install warnings on newer Typer releases.
Validation¶
Before release, the focused release checks passed:
flake8on the changed simulator, Jetson, research, CLI, and test filespytest tests/core/test_safety_config.py tests/core/test_numeric_guards.py tests/algorithm/test_clinical_baseline.py tests/research/test_control_training.py tests/test_jetson_endurance.py tests/test_cli_onboarding.py -q48 passedpytest tests/research/test_control_training.py tests/test_jetson_endurance.py -q19 passed
A 24-hour accelerated Jetson smoke run also completed 288 steps and produced the new physiology and local-AI research artifacts.
Upgrade¶
For normal SDK and reporting work:
python -m pip install -U "iints-sdk-python35[full,mdmp]==1.5.7"
For local AI training features, include the research extra:
python -m pip install -U "iints-sdk-python35[full,mdmp,research]==1.5.7"
For edge/Jetson installs where disk space matters:
python -m pip install -U "iints-sdk-python35[edge,mdmp]==1.5.7"
Practical takeaway¶
Use iints jetson endurance start --wall-clock to acquire the run, then use iints research local-ai-lab to turn one or more completed runs into datasets, local models, and held-out evaluation evidence.