Responsible AI & Clinical Safety

AI boundaries, deterministic engine positioning, human-in-the-loop requirements, and auditability guarantees for the Clinical Temporal Decision Engine.

Deterministic Engine — AI as Orchestration Only

The Clinical Temporal Decision Engine is a deterministic system. The engine itself does not use generative AI, large language models (LLMs), or machine learning for clinical reasoning.

MCP Agent Surface

Exposes pipeline tools and FAQ knowledge base via Model Context Protocol for integration with agent frameworks. Navigation and orchestration only.

FAQ Assistant

Deterministic keyword and category-based search over a curated knowledge base. Not a chatbot. Not generative. Every result is a pre-authored, reviewed entry.

Pipeline Execution

Rule-based guideline evaluation with versioned, auditable execution. No ML-based clinical inference. Every output is reproducible.

The deterministic engine is the source of truth. AI/MCP layers navigate and orchestrate access to the engine — they do not perform clinical reasoning.

AI / ML Boundaries

What AI Does

  • Navigate pipeline tools via MCP (tool discovery and invocation)
  • Search FAQ knowledge base (deterministic keyword matching)
  • Orchestrate multi-step workflows (e.g., “run guideline X against bundle Y”)
  • Surface audit trails and documentation (retrieval, not generation)

What AI Does NOT Do

  • Diagnose medical conditions
  • Recommend treatments or medications
  • Perform differential diagnosis
  • Generate clinical notes or summaries from patient data
  • Autonomously make or recommend clinical decisions
  • Replace or override clinical guidelines
  • Process or reason about real patient data (synthetic data only in current state)

Human-in-the-Loop

Clinical decision support requires human oversight:

  • All pipeline outputs are decision support, not clinical directives.
  • Every risk score is traceable to input data and guideline version for human review.
  • Quality gate failures produce structured reports requiring human attention.
  • Care gap outputs flag potential issues for clinical review — they do not dictate clinical action.
  • Audit trails support independent verification of every computation.

A qualified healthcare professional must review all outputs before any clinical or operational action. The system does not operate autonomously in any clinical decision path.

Auditability and Reproducibility

Every output from the engine is:

Traceable

Input → normalization → quality gate → guideline execution → score

Reproducible

Same input + same guideline version = same result, always

Versioned

Guideline version and execution timestamp recorded

Auditable

Full provenance chain available for every computation

This is not an AI assurance claim. It is a deterministic engineering guarantee.

Synthetic Data Readiness

Current system demonstrations, testing, and development use Synthea-generated synthetic data exclusively:

  • No real patient data is processed in any environment.
  • All screenshots reflect synthetic data with visible labels.
  • Synthetic data is generated from published Synthea modules and does not contain real patient identifiers, MRNs, or PHI.
  • Transition to real data handling requires compliance program implementation and legal review — this has not yet occurred.

Demonstrations use synthetic data only. No real patient data is used in development, testing, or demonstration.

Not a medical device. Does not diagnose, treat, or prescribe.

Not HIPAA, FDA, SOC2, or ISO certified. Compliance program is planned.