Customer operations need more than a persuasive answer.
Live service decisions combine changing conversation state, commercial policy, protected facts, delivery constraints and asynchronous follow-up. ARMIE treats these as explicit system responsibilities. The model can help interpret or explain; it does not silently own eligibility, operational truth or lifecycle state.
Same discipline, different engineering questions.
Which state is allowed to become truth?
A FastAPI reference slice separates transcript revision, versioned snapshots, protected facts, business outcome and delivery lifecycle. Stale asynchronous work cannot overwrite newer state.
- Browser Workbench and canonical conversation state
- Monotonic UI sequencing and permission rechecks
- Post-call lineage, idempotency and delete tombstones
- SQLite default plus opt-in PostgreSQL persistence contract
- Docker / EC2 demo path with CloudWatch and SSM
Which commercial action is policy-authorized?
A governed decision workflow combines customer context, deterministic eligibility, synthetic guidance retrieval, bounded narration, validation and human decision support.
- ML predicts; deterministic policy authorizes
- Retrieval grounds; the LLM explains
- Validator and fallback protect the output boundary
- Container Apps, Azure OpenAI, Speech, Language, Event Hubs and Redis validated in bounded paths
- Workbench exposes decision evidence and fallback behavior
Versioned state prevents stale work from becoming current truth.
The PostgreSQL adapter demonstrates a persistence contract and transaction behavior. It does not establish RDS deployment, live migration, multi-instance safety or recovery of live calls.
Policy remains deterministic while narration stays replaceable.
Eligibility and action selection remain governed by explicit policy.
The model explains an authorized decision rather than inventing one.
Selected Azure paths were measured directly; the results are engineering evidence, not an SLA.
Guidance lookup is a labeled local synthetic stand-in; Azure AI Search is not claimed as the live path.
These are two independently built reference implementations using synthetic scenarios. The page does not claim customer deployment, production scale, contact-center integration, cross-cloud failover, commercial outcomes or one unified multi-cloud product.
Implemented systems, with production questions still open.
The AWS system demonstrates governed live-conversation state and asynchronous enrichment. The Azure system demonstrates governed commercial decision support and bounded AI narration. Production data integration, security operations, tenancy, scale, recovery objectives and customer-specific policy remain future delivery concerns rather than current public claims.