The GPT 5.6 model path is connected to the OpenAI Platform and verified in the current Azure Dev runtime.
Public evidence covers implemented engineering paths, synthetic data and repository-backed verification.
Why this exists
This project explores how a local healthcare reasoning pipeline can become an observable, authenticated, durable and cloud-operable AI system without moving core context, retrieval, validation or workflow semantics into a model vendor. It is a reference architecture for engineering patterns, not a healthcare chatbot, clinical product or source of medical advice.
System architecture
Engineering evolution
Durable workflow & reliability
Durable Functions coordinates long-running execution, bounded retry, timers, cancellation and replay/recovery.
Cosmos-backed ETag/CAS concurrency prevents conflicting terminal outcomes.
Stable business run_id and idempotency-key semantics prevent accidental duplicate workflow creation.
Success, timeout and cancellation are competing terminal events; the Workbench preserves backend truth.
Persistent state & evidence
Run identity, lifecycle, requested/executed reasoning mode, validation and persistence status, orchestration correlation, attempts and terminal ownership.
Retrieved evidence, structured traces, result artifacts and evaluation/safety artifacts remain durable and auditable.
Function compute remains replaceable while operational state and evidence remain durable and auditable.
Identity & security architecture
Implemented Azure Dev boundaries include Microsoft Entra single-tenant identities, Easy Auth-protected run endpoints, a browser MSAL path for the Workbench, system-assigned Managed Identity and narrow Cosmos/ADLS data-plane access. The project uses synthetic/demo data only; no HIPAA, GDPR, PHI or other regulatory certification is claimed.
Safety & evaluation
Safety checks execute as an explicit boundary after reasoning rather than being delegated to the reasoning model.
Rule-based reasoning provides repeatable CI and characterization behavior.
Retrieved evidence, reasoning provenance and execution traces remain inspectable.
Synthetic data and repository-backed scenarios support engineering evaluation, not clinical efficacy or safety benchmarking.
An operator and evaluation surface for the runtime.
The React/Vite Workbench is a projection of backend runtime truth, not a second implementation of healthcare business logic. It supports local FastAPI and authenticated Azure Dev backends, synchronous and Durable execution, run/orchestration identifiers, lifecycle projection, attempt and persistence state, cancellation control and explicit terminal outcomes.
Model provider strategy
ruleDeterministic baseline / CIlocal_llmLocal Ollama / Qwen experimentationopenaiOpenAI Platform · GPT 5.6 integration validatedConfirmed capabilities
Healthcare context, retrieval, reasoning integration and validation remain in the ARMIE runtime.
Inspectable local and authenticated Azure Dev execution surface.
Serverless authenticated API/runtime boundary for synthetic Azure Dev execution.
Recoverable workflow execution with retry, timeout and cancellation semantics.
Cosmos DB stores queryable run and terminal-state truth.
ADLS Gen2 stores traces, evidence and larger run artifacts.
Entra, Easy Auth and Managed Identity define user/service boundaries.
Idempotency, concurrency and terminal ownership are explicit system semantics.
Safety remains a separate validation boundary after reasoning.
Application Insights and structured traces expose runtime behavior.
Rule and local LLM paths remain available while the cloud provider remains replaceable.
Current scope and maturity
This is an ARMIE AI synthetic, non-clinical engineering reference system. Azure Dev infrastructure, authenticated execution, Durable Functions orchestration, Cosmos DB and ADLS persistence, Workbench integration, Managed Identity, structured observability and the OpenAI Platform GPT 5.6 path have been implemented and validated with synthetic data.
It is not a production healthcare deployment, hospital system, diagnostic or prescribing product, regulated medical device or compliance-certified application. No PHI processing, production SLA or clinical safety benchmark is claimed.