Status: v0.1
Objective#
Design the operating system and SOPs for a ProxI-enabled CMS GUIDE Participant that maximizes safe automation, minimizes expensive human effort, satisfies GUIDE care-delivery and documentation requirements, and creates reproducible audit evidence.
Product baseline#
ProxI Base Camp provides:
- A voice-based caregiver digital twin that interacts with the person living with dementia.
- A caregiver application for status, feedback, preferences, and intervention.
- Continuous patient/caregiver touchpoints and event intake.
- Structured Care Space storage with provenance, staleness, candidate facts, confirmed facts, decisions, and versioned plans.
- Integrations with clinical, pharmacy, claims, messaging, and device data.
- Natural-language delivery of content selected by deterministic logic.
Design objectives#
- Automate data acquisition, normalization, comparison, routing, reminders, retries, and evidence assembly.
- Reduce human work to explicit exceptions and authority gates.
- Place each exception with the lowest-cost role legally and operationally permitted to resolve it.
- Keep Philippine personnel in administrative pursuit and preparation roles unless a later written determination expands that boundary.
- Make stable patient months require effectively no PCP action.
- Preserve complete provenance and replayability for every decision and action.
- Produce a staffing model from rule incidence and measured minutes rather than blended intuition.
Non-goals#
- Autonomous diagnosis, prescribing, deprescribing, treatment recommendation, or clinical triage by an LLM.
- Treating AI interaction as the required human care-navigator contact.
- Using silent computation as evidence that a GUIDE service was furnished.
- Requiring outside PCPs to install or log into a new ProxI portal.
- Closing referrals without returned documentation and care-plan reconciliation.
- Treating prescribed, dispensed, possessed, and actually taken medications as one state.
Deliverables#
- Pillar-specific SOPs and state machines.
- Cross-pillar role and escalation matrix.
- Edge-case register.
- Deterministic rule catalog.
- Evidence and closure definitions.
- Test-case library.
- Staffing and cost model driven by observed exception incidence.
- Decision log separating settled assumptions from unresolved choices.
Acceptance criteria#
The architecture is not complete until:
- Every state transition has deterministic prerequisites.
- Every unresolved state has an owner and SLA.
- Every human role has explicit permitted and prohibited actions.
- Every closure state has objective evidence.
- Every material rule has positive, negative, ambiguous, and failure-path tests.
- No action-capable interface accepts raw LLM output as a command.