Skip to content

Local AI Safety Gates

The local-AI pipeline is intentionally critical. A controller can train successfully and still be blocked from research promotion.

What Gets Checked

The training gate reviews:

  • minimum training rows;
  • maximum teacher insulin label;
  • unsafe insulin proposals during hypoglycemia;
  • proposals above 5 U;
  • whether the training target came from reference_teacher_insulin_units;
  • whether deterministic supervisor limits are still required.

The held-out closed-loop gate reviews every candidate against clinical_baseline:

  • no early terminations;
  • completion percentage;
  • time below 54 mg/dL regression;
  • time below 70 mg/dL regression;
  • supervisor-intervention regression;
  • large TIR drop.

Why This Matters

For diabetes technology research, a model that improves average glucose but increases severe hypoglycemia is not an improvement. The SDK therefore treats hypo regression and safety-supervisor burden as blocking issues.

Output

iints research local-ai-lab now writes:

  • training_safety_gate in LOCAL_AI_RESEARCH_SUMMARY.json;
  • closed_loop_evaluation.safety_gate when evaluation is enabled;
  • a report section explaining why a local model is blocked, review-only, or research-ready.

Boundary

These gates are pre-clinical research gates. They do not approve a model for pump control or patient dosing.