safe-diabetes-t2d-adolescent-cup

CUP: t2d / adolescent

September–October 2026 simulator-reset rerun. This release contains three policy checkpoints (seeds 100, 101, 102, endpoint epoch 976) and 18 matching direct CGM predictors: three policies × two action modes × three predictor seeds (3101, 3102, 3103). Policy weights include their frozen observation normalizer. The policy uses 14 observations and independent categorical bolus/meal levels 0–4. A requested or simulator-accepted level is not a delivered physical dose.

Policies were trained online on this cohort's patient #001, with a one-day horizon. Predictors used only seen #001 train/validation/test transition splits. Predictor weights are CausalMLP, prediction_mode=direct. The legacy artifact schema name ba_node_predictor does not mean these are BA-NODE models.

Seven-day unseen-patient results

Patients #002–#010, reset/action seeds 83000, 83001, 83002; both action modes are reported separately. Each mean first averages resets and patients, then predictor seeds within each policy when applicable, then the three policy means equally. ± is sample SD across the three policy means. None/static use 81 physical episodes per mode; predictive uses 243. TIR counts CGM 70–180 mg/dL over the requested horizon; early endings remain in its denominator.

The selected absolute penalty comes from 1/5/10/15/20 on separate seen #001 one-day validation, never the unseen test. no_eligible_penalty is a completed scientific NA outcome. Selected10 reuses fixed10. These are existing authenticated audit-summary results, not new reruns. Shield benefits vary by cohort, policy and metric. TIR alone does not establish safety; early endings, coverage and low glucose outcomes accompany it. The six-algorithm export does not include its separate archive-wide native reward/cost supplement.

Download and loading

Install the dependencies for the public GlucoAlg loading revision and GlucoSim. This is a custom PyTorch/GlucoAlg policy, with these relative paths:

import json
from pathlib import Path
from huggingface_hub import snapshot_download
from glucoalg.evaluation import create_diabetes_env, load_model

root = Path(snapshot_download("safe-diabetes-benchmark/safe-diabetes-t2d-adolescent-cup"))
release = root / "releases/rerun-20261008"
config = json.loads((release / "config/seed100/config.json").read_text())
env = create_diabetes_env("t2d", "adolescent#002", seed=83000, horizon_days=7)
actor, env, normalizer = load_model(
    release / "checkpoints/seed100/epoch-976.pt", config, env,
    shield_type="none",
)

load_model(..., shield_type="rule_based", logit_penalty=10) selects the static shield. Predictive shielding requires the explicit adapter and dynamics rollout workflow described in predictive-shield.md, not a load_model predictive flag:

from glucoalg.dynamics.model import load_predictor
adapter = load_predictor(
    release / "predictors/seed100/deterministic/model3101", device="cpu"
)

Match policy seed, action mode and predictor seed as specified in release.json. Normalize policy observations using the saved normalizer, and pass raw observations to shield forecasting. Config and predictor artifacts retain exact training bytes and hashes; their old absolute paths are provenance, not portable runtime locations. Predictor triplets must stay together unchanged.

Provenance and license

The loading revision differs from the frozen training sources. Original-policy training and repaired-PCPO training sources are identified separately by the release manifest. Simulator commit: ddd674e35a32a3a6406a0c2d5e7bf3c72f56dfec. Local package validation checks CPU loading and integrity; it is not independent scientific or clinical validation. The models and associated documentation use MIT; this does not change source-code licenses.

The January 2026 files are replaced on the default branch. The existing old revision is preserved under legacy-20260131; it is distinct from this rerun. The accompanying synthetic transition dataset is predictor training data, not offline policy-training data. These simulator research artifacts are not validated clinical dosing systems.

Related paper: Safety Generalization Under Distribution Shift in Safe Reinforcement Learning: A Diabetes Testbed, Minjae Kwon, Josephine Lamp and Lu Feng. The May 2026 paper version predates these September–October rerun numbers.