Healthcare intelligence architecture that treats models as interfaces, not the whole system.

meta-harness-low-risk-code-review

• Tools: read_only_code_inspector:allowed, schema_validator:allowed, audit_recorder:allowed
• No PHI confirmed: yes
• No external side effects: yes
• Audit: scrimed-intel-27f15c5e

meta-harness-clinical-review-packet

• Tools: synthetic_context_loader:allowed, schema_validator:allowed, policy_gate:allowed, evidence_binder:allowed, human_approval_queue:allowed, audit_recorder:allowed
• No PHI confirmed: yes
• No external side effects: yes
• Audit: scrimed-intel-cea21115

meta-harness-blocked-protected-action

• Tools: payer_submission:blocked, patient_outreach:blocked, ehr_writeback:blocked, raw_payload_logger:blocked
• No PHI confirmed: yes
• No external side effects: yes
• Audit: scrimed-intel-7e70933a

patient matching

• Traceability: 70
• Minimum passing score: 80
• Candidate sources: preindex-fhir-synthetic-care-summary
• Human review required: yes
• Audit: scrimed-intel-d11177b7

document similarity

• Traceability: 70
• Minimum passing score: 80
• Candidate sources: preindex-payer-policy-synthetic-afib, preindex-referral-synthetic-delay, preindex-trial-synthetic-protocol
• Human review required: yes
• Audit: scrimed-intel-68d78a90

clinical retrieval

• Traceability: 70
• Minimum passing score: 80
• Candidate sources: preindex-fhir-synthetic-care-summary, preindex-guideline-synthetic-care-gap, preindex-trial-synthetic-protocol
• Human review required: yes
• Audit: scrimed-intel-ec192c55

payer policy lookup

• Traceability: 70
• Minimum passing score: 80
• Candidate sources: preindex-payer-policy-synthetic-afib, preindex-rcm-synthetic-denial-playbook
• Human review required: yes
• Audit: scrimed-intel-238a7ad6

recommendation search

• Traceability: 70
• Minimum passing score: 80
• Candidate sources: preindex-fhir-synthetic-care-summary, preindex-referral-synthetic-delay, preindex-guideline-synthetic-care-gap, preindex-rcm-synthetic-denial-playbook
• Human review required: yes
• Audit: scrimed-intel-94cd4457

deid-pdf-synthetic-intake

• Targets: browser, mac
• External inference allowed: no
• Raw payload stored: no
• Human verification required: yes
• Audit: scrimed-intel-c4a02514

deid-scan-synthetic-referral

• Targets: mac, iphone
• External inference allowed: no
• Raw payload stored: no
• Human verification required: yes
• Audit: scrimed-intel-43c78ea5

deid-image-synthetic-card

• Targets: browser, iphone
• External inference allowed: no
• Raw payload stored: no
• Human verification required: yes
• Audit: scrimed-intel-b9bde5a3

deid-hl7-v2-synthetic-adt

• Targets: mac
• External inference allowed: no
• Raw payload stored: no
• Human verification required: yes
• Audit: scrimed-intel-e4cf5f1e

deid-cda-synthetic-summary

• Targets: browser, mac
• External inference allowed: no
• Raw payload stored: no
• Human verification required: yes
• Audit: scrimed-intel-ec2c8e0b

deid-fhir-synthetic-bundle

• Targets: browser, mac
• External inference allowed: no
• Raw payload stored: no
• Human verification required: yes
• Audit: scrimed-intel-a370ea14

deid-csv-synthetic-roster

• Targets: browser, mac
• External inference allowed: no
• Raw payload stored: no
• Human verification required: yes
• Audit: scrimed-intel-d1aa70eb

deid-ndjson-synthetic-export

• Targets: mac
• External inference allowed: no
• Raw payload stored: no
• Human verification required: yes
• Audit: scrimed-intel-efa2f0ab

deid-chat-log-synthetic-support

• Targets: browser, mac, iphone
• External inference allowed: no
• Raw payload stored: no
• Human verification required: yes
• Audit: scrimed-intel-bb21a697

doc-auth-afib-ablation-synthetic-gap

• Prior-auth risk: missing symptom language, missing functional status, missing visit timing, missing recent visit note, missing reviewer attestation, human reviewer attestation incomplete
• Automation: blocked
• Human review required: yes
• Payer submission allowed: no
• Audit: scrimed-intel-40ecb0ad

doc-auth-imaging-synthetic-review-ready

• Prior-auth risk: low synthetic risk
• Automation: recommendation_only
• Human review required: yes
• Payer submission allowed: no
• Audit: scrimed-intel-13d48030

Omnigent-style Meta-Harness

• Lane: single orchestrator contract, shared session context, agent permission manifest, policy gate, human approval checkpoint
• Outputs: orchestration plan, tool access decision, approval route, audit hash, blocked action list
• Oversight: Human approval remains required before any protected clinical, payer, connector, outreach, or system-of-record action.
• Boundaries: no live PHI, no autonomous clinical authority, no direct database or connector action by models

Documentation-Before-Authorization Engine

• Lane: payer-policy checklist schema, documentation gap detector, medical necessity phrase map, visit timing validator, pre-submission risk packet
• Outputs: documentation gap list, prior-auth risk score, reviewer packet, missing evidence map, submission blocked status
• Oversight: RCM or clinical reviewer must approve any payer-facing packet; SCRIMED does not submit prior authorizations or claims.
• Boundaries: no payer submission, no billing submission, no medical necessity final determination

Edge Clinical Trial Evidence Layer

• Lane: edge evidence envelope, site/device provenance, tamper-evident hash, offline sync status, trial monitor review queue
• Outputs: evidence envelope, provenance hash, validation status, sync readiness, monitor review flag
• Oversight: Trial monitor, investigator, or sponsor-designated reviewer must validate before operational or regulatory use.
• Boundaries: no live participant data, no regulatory submission, no clinical trial enrollment action

On-Device De-Identification

• Lane: document-type detector, PHI span scanner, redaction manifest, local-first processing policy, de-identification audit summary
• Outputs: redaction manifest, document class, PHI risk score, human verification flag, blocked upload reason
• Oversight: Human verification remains required before any de-identified artifact is used beyond synthetic demo mode.
• Boundaries: no raw PHI upload, no raw connector payload logging, no external inference without explicit authorization

Clinical AI Benchmark Lab

• Lane: specialty benchmark suites, physician grader rubric, source-quality scoring, verifiability checks, completeness and utility scoring
• Outputs: clinical readiness score, source quality score, verifiability score, utility score, review requirement
• Oversight: Qualified physician graders or designated clinical reviewers remain the authority for benchmark disposition.
• Boundaries: no diagnosis, no treatment recommendation, no clinical validation claim

Automation Orchestrator

• Lane: workflow state machine, tool invocation registry, approval gate, retry and fallback policy, audit event stream
• Outputs: workflow plan, safe tool decision, approval checkpoint, retry decision, audit trace
• Oversight: Automation can recommend or prepare actions, but protected healthcare actions remain blocked or approval-gated.
• Boundaries: no autonomous outreach, no EHR writeback, no payer or billing submission

Pre-Indexed Intelligence

• Lane: ingest-time parser, structure-preserving chunker, provenance index, retrieval evaluation set, grounding quality report
• Outputs: enriched index record, provenance chain, retrieval score, grounding report, staleness flag
• Oversight: Governance reviewer approves production indexes, source scope, retention, and connector activation.
• Boundaries: no raw schema exposure to agents, no production connector approval, no unreviewed source ingestion

AI Medical Education Layer

• Lane: clinician training track, skill assessment rubric, AI literacy modules, feedback capture, personalized learning plan
• Outputs: learning plan, skill assessment, feedback packet, readiness milestone, non-certification disclaimer
• Oversight: Education leaders and clinical governance reviewers approve curriculum and assessment language before external use.
• Boundaries: no credentialing claim, no certification claim, no clinical competency authorization

Treat LLMs as the interface layer, not the whole system.

• Architecture: Separate model interface adapters from workflow state machines., Route all protected actions through policy, schema, evidence, and human-review services., Keep model output as draft or recommendation until validated.
• Tracks: Model adapter interface, Workflow orchestrator, Policy gate, Evidence binding, Human review queue
• Validation: Route contract tests prove models do not write directly to systems of record., Generated outputs must pass structured schema checks., Human-review flags must remain true for protected healthcare workflows.
• Next: Add interface-layer tags to model routes so every AI output declares draft, recommendation, or metadata-only status.

Add world-model/context layers for messy healthcare data.

• Architecture: Create context layers for patient journey state, clinical workflow state, payer state, RCM state, access state, geography, time, capacity, and evidence freshness., Add uncertainty, missingness, conflict, source, and timestamp fields to context packets., Compress context before model use while preserving provenance.
• Tracks: Context packet schema, World-model state registry, Uncertainty labels, Source freshness checks, Conflict detection
• Validation: Synthetic messy-data fixtures test missing, conflicting, stale, temporal, and unit-inconsistent inputs., Context packets must expose uncertainty and missingness., Reviewer queues receive escalation when context is insufficient.
• Next: Create synthetic world-model fixtures for time-series, geography, physical constraints, payer rules, workflow state, and patient journey state.

Build active ontology + semantic graph for clinical, payer, RCM, patient-access, and operations logic.

• Architecture: Create ontology domains for clinical, payer, RCM, patient access, operations, workforce, resource, and governance logic., Bind semantic graph nodes to evidence, policy, workflow state, and reviewer ownership., Require graph-derived constraints before model execution.
• Tracks: Ontology registry, Semantic graph node schema, Relation vocabulary, Evidence lineage, Policy constraint resolver
• Validation: Graph nodes must have type, owner, evidence requirement, and safety boundary., Edges must be constrained to known relationship types., Contradictory graph paths must route to human review.
• Next: Extend the TrustOps semantic graph with ontology node types for payer rules, RCM denial logic, access queues, workforce capacity, and resource constraints.

Store agent reasoning traces, audit logs, and decisions as long-term memory.

• Architecture: Persist decision trace metadata, cited rationale summaries, inputs hashes, evidence refs, policy refs, model route, tool calls, and reviewer disposition., Separate human-readable rationale summaries from private model chain-of-thought., Tie long-term memory to retention, deletion, tenant, residency, and review controls.
• Tracks: Decision memory schema, Audit event ledger, Evidence envelope hash, Reviewer disposition, Retention and deletion policy
• Validation: No hidden model chain-of-thought is required or exposed., Memory records must be metadata-only unless future PHI approval exists., Every persisted decision must include evidence refs, policy refs, and reviewer status.
• Next: Map TrustOps review packets and execution-attempt durable envelopes into a shared long-term decision-memory contract.

Add dynamic context injection: deep reasoning at session start, skill/module listing every turn, and task reminders updated every turn.

• Architecture: Create a context-injection manifest for session start, every turn, and task completion., Run deep planning summaries at session start without exposing hidden chain-of-thought., List relevant skills/modules every turn and refresh task reminders as state changes.
• Tracks: Session context primer, Per-turn skill/module manifest, Task reminder ledger, Lazy capability loading, Context compression
• Validation: Context manifests must cite selected modules and omitted modules., Task reminders must be versioned and updated after each workflow state change., Prompt payloads must exclude secrets, PHI, and irrelevant tools.
• Next: Create a per-turn context manifest schema with selected modules, skill list, active reminders, omitted context, safety boundaries, and evidence refs.

Avoid the self-correction trap: never trust model self-verification alone; validate with schemas, evidence, external data, rules, and human review.

• Architecture: Require schema validation for structured outputs., Bind claims to evidence cards and source freshness., Use deterministic rule checks before release., Escalate missing, conflicting, or high-risk outputs to human review.
• Tracks: Structured-output validator, Evidence verifier, Rule engine, Human-review gate, Regression benchmark
• Validation: Every generated brief must pass schema fidelity checks., Evidence and source attribution must be present for clinical or operational claims., Human-review status must remain unresolved until qualified disposition.
• Next: Promote schema, evidence, rules, external-data hooks, and reviewer disposition into a release gate for all TrustOps and module briefs.

Add workforce/talent module for healthcare hiring, onboarding, vacancy-risk, and labor-cost savings.

• Architecture: Create workforce demand, vacancy risk, onboarding readiness, credential checklist, training state, and labor-cost model objects., Connect workforce signals to access, scheduling, referral, and operations bottlenecks., Keep employment, HR, legal, and finance claims review-gated.
• Tracks: Workforce capacity model, Vacancy-risk signal, Onboarding checklist, Labor-cost savings model, Hiring readiness packet
• Validation: Use synthetic staffing scenarios only., Cost-savings claims require assumptions, ranges, evidence, and finance review., No hiring, employment, legal, or payroll action is automated.
• Next: Add workforce/talent registry entries to TrustOps with synthetic vacancy-risk and onboarding-readiness signals.

Add project/resource management module tracking compute, storage, quota, model usage, pipeline cost, and agent workload.

• Architecture: Track compute, storage, quota, model usage, pipeline cost, agent workload, review queue load, and tenant capacity., Connect resource signals to cost guardrails, model routing, deployment readiness, and sales/package margin controls., Create budget thresholds and recommendation-only mitigation packets.
• Tracks: Resource usage ledger, Model cost telemetry, Pipeline cost estimator, Agent workload queue, Budget guardrail signal
• Validation: Synthetic usage scenarios test quota exhaustion, model cost spikes, pipeline delays, and overloaded agent queues., Cost estimates must state assumptions and confidence., Mitigations are recommendation-only until approved.
• Next: Bind model usage, pipeline cost, quota, storage, and agent workload to the TrustOps Signal Engine as synthetic cost-risk signals.

Build toward healthcare world models: time-series, geography, physical constraints, clinical workflow state, payer rules, and patient journey state.

• Architecture: Create world-model layers for time-series trends, geography, physical capacity, clinical state, payer rules, patient journey, workforce, and resource constraints., Represent temporal ordering, dependencies, uncertainty, state transitions, and blocked actions., Use world-model outputs to constrain workflow selection and escalation.
• Tracks: Time-series state layer, Geography and facility layer, Physical constraints layer, Clinical workflow state layer, Payer rules layer, Patient journey state layer
• Validation: Synthetic scenarios test impossible timing, geography mismatch, capacity conflicts, payer-rule conflicts, and incomplete journey state., World-model conflicts must block automation and route to review., Outputs must expose limitations and confidence.
• Next: Create the first synthetic world-model test suite covering temporal order, facility geography, physical capacity, payer rules, and journey state.

Add benchmark layer for structured outputs, schema fidelity, reasoning validity, and operational accuracy.

• Architecture: Create benchmark suites for schema fidelity, evidence grounding, reasoning validity, semantic graph consistency, workflow accuracy, and operational impact., Tie benchmark results to release gates and TrustOps scores., Track regressions by module, prompt, model route, and workflow version.
• Tracks: Benchmark Studio, ClinicalBench, TrustOps score integration, Regression report, Release gate
• Validation: Structured outputs must match schemas., Reasoning summaries must be evidence-grounded and rule-consistent., Operational outputs must match expected workflow state and owner routing.
• Next: Add benchmark dimensions to the TrustOps registry and require benchmark status in every build-roadmap release summary.

Dynamic Context Injection Engine

• Inputs: task request, module registry, skill registry, policy gate, active reminders, evidence refs
• Outputs: context manifest, selected module list, task reminder update, omitted context log, safety boundary note
• Controls: no secrets, no PHI, lazy capability loading, context compression, permission-aware tool listing
• Blocked: tool permission grant, secret exposure, PHI injection, protected action approval

Active Ontology + Semantic Graph

• Inputs: ontology registry, policy refs, evidence refs, workflow state, synthetic events
• Outputs: semantic graph, constraint map, conflict report, evidence lineage, owner routing
• Controls: typed nodes, known relation vocabulary, evidence required, human review on conflict
• Blocked: untyped free-form graph mutation, source-less claims, autonomous system-of-record action

Decision Memory Ledger

• Inputs: execution attempt envelope, TrustOps review packet, audit event, reviewer disposition, benchmark result
• Outputs: decision memory record, evidence envelope hash, review history, regression trigger, retention event
• Controls: metadata-only by default, retention policy, deletion policy, tenant scope, no hidden chain-of-thought
• Blocked: PHI memory without approval, hidden chain-of-thought disclosure, cross-tenant replay

Workforce / Talent Intelligence

• Inputs: synthetic staffing scenarios, role catalog, onboarding checklist, capacity assumptions, labor-cost assumptions
• Outputs: vacancy-risk signal, onboarding-readiness packet, staffing capacity estimate, labor-cost savings range
• Controls: synthetic-only, finance review, HR/legal review, assumption disclosure, no payroll mutation
• Blocked: hiring decision, employment advice, payroll action, credentialing approval, labor-law advice

Project / Resource Management Intelligence

• Inputs: synthetic usage events, model route telemetry, pipeline estimates, quota thresholds, agent queue load
• Outputs: cost-risk signal, quota-risk signal, agent workload report, pipeline cost estimate, margin protection packet
• Controls: budget guardrails, assumption disclosure, human approval, no infrastructure mutation, no financial guarantees
• Blocked: cloud mutation, billing mutation, service shutdown, secret rotation, financial guarantee

Operational Benchmark Layer

• Inputs: synthetic scenarios, expected schemas, rules, evidence cards, workflow expected state, reviewer rubric
• Outputs: benchmark report, schema-fidelity score, reasoning-validity score, operational-accuracy score, release gate
• Controls: schema validation, evidence validation, rule checks, external data hooks, human review
• Blocked: leaderboard-only claims, self-verification-only release, clinical validation claim

Time-Series Layer

• event time, workflow age, trend window, freshness, sequence validity
• Validation: synthetic event stream, timestamp rules, freshness checks, temporal contradiction tests
• Blocked until: Live patient timelines require PHI approval, consent/data-use review, retention rules, and customer authorization.

Geography Layer

• facility location, service area, region, travel constraint, jurisdiction
• Validation: synthetic facility map, region policy refs, routing constraints, jurisdiction checks
• Blocked until: Real location or patient travel data requires approved privacy and customer governance.

Physical Constraints Layer

• capacity, resource availability, equipment state, queue load, slot feasibility
• Validation: synthetic capacity fixtures, queue simulations, constraint solver checks, operations review
• Blocked until: Production scheduling, staffing, or equipment actions require customer approval and connector review.

Clinical Workflow State Layer

• workflow status, review state, risk level, evidence sufficiency, blocked action
• Validation: ClinicalBench, reviewer rubric, evidence cards, policy gate
• Blocked until: Live clinical workflows require clinical governance, PHI controls, and qualified human review.

Payer Rules Layer

• policy version, criteria match, documentation gap, denial reason, manual verification state
• Validation: synthetic payer-policy fixture, rule engine, RCM reviewer rubric, source freshness
• Blocked until: Real payer submissions, appeals, or claims require customer, payer, legal, and compliance approval.

Patient Journey State Layer

• journey milestone, handoff, care gap, access state, continuity risk
• Validation: synthetic journey fixture, care-gap benchmark, human review, policy boundary
• Blocked until: Live patient journey memory requires PHI, consent, retention, deletion, and customer approval.

scrimed-context-manifest-synthetic-pre-agent-run-v1

• Safety decision: allowed
• Prompt boundary: Prompt payload may include synthetic task objective, selected module summaries, skill/module listing every turn, active task reminders, validators, evidence refs, omitted-context log, and safety boundaries. It must exclude PHI, secrets, credentials, hidden chain-of-thought, irrelevant tools, and production connector payloads.
• Memory: Store concise rationale summaries and evidence refs only; do not request, expose, or persist hidden model chain-of-thought.
• Audit hash: synthetic-57c2bdb3

Lazy capability loadout

• Loaded: dynamic-context-injection-engine, operational-benchmark-layer, workflow-planner, governance-compliance, secure-middleware-gateway, schema-validator, safety-policy-validator
• Deferred: researchops, clinicalbench, semantic-intelligence-graph, signal-detection, self-healing-operations
• Blocked: live-phi-context, production-ehr-connector, production-payer-submission, patient-outreach-tools, billing-submission-tools
• Block the run when PHI, secrets, production connector payloads, missing validators, or protected actions appear.

reminder-no-phi

• Updated: 2026-06-30T00:00:00.000Z
• Required before next run: yes

reminder-validate-not-self-correct

• Updated: 2026-06-30T00:00:00.000Z
• Required before next run: yes

reminder-context-minimization

• Updated: 2026-06-30T00:00:00.000Z
• Required before next run: yes

pre-agent-run-context-manifest-ready

selected-modules-listed-every-turn

task-reminders-versioned

omitted-context-protects-phi-secrets-connectors-and-hidden-chain-of-thought

validators-avoid-self-correction-trap

manifest-cannot-grant-tool-access

memory-plan-metadata-only-no-hidden-cot

safety-governance-pass

Agent Runtime context bridge

• Required before every synthetic agent run: yes
• Selected modules: dynamic-context-injection-engine, operational-benchmark-layer, workflow-planner, governance-compliance, secure-middleware-gateway
• Validators: schema-validator, evidence-validator, rule-validator, human-review-validator, safety-policy-validator, operational-benchmark-validator
• Block agent planning when context manifest is missing, invalid, policy-blocked, PHI-bearing, secret-bearing, or missing human-review validators.

Stored-vector lookup backend RPCs

• RPCs: register_scrimed_synthetic_stored_vector, scrimed_match_stored_vector, scrimed_search_similar_documents, scrimed_search_clinical_evidence, scrimed_search_payer_policy, scrimed_search_recommendation_memory
• Controls: AAL2 governance session required, server runtime token required, tenant-admin/pilot-lead registration only, tenant-admin/pilot-lead/reviewer lookup only
• Run the authenticated stored-vector RPC smoke with a short-lived AAL2 tenant-admin or pilot-lead token after the app route is deployed and the protected server token is configured.

patient-matching

• Reduces client round trips, embedding serialization, network transfer, and tail-latency variance.
• Synthetic/de-identified matching only; no live patient identity resolution or production patient matching authority.
• Requires pgvector index review, RLS/tenant predicate, audit event, no service-role browser exposure, and PHI approval before live use.

document-similarity

• Keeps vector math close to the index and avoids large embedding payload serialization.
• No live chart documents; public, synthetic, or approved de-identified document evidence only.
• Requires source attribution, document retention policy, RAG poisoning checks, and reviewer-governed ingestion.

clinical-retrieval

• Cuts prompt bloat and improves grounding by returning ranked evidence cards instead of raw context dumps.
• Research/demo only; not diagnosis, treatment, prescribing, or live patient care.
• Requires clinical evidence steward, freshness policy, citation verification, and qualified reviewer workflow.

payer-policy-lookup

• Reduces policy payload movement and keeps policy-version constraints deterministic before model synthesis.
• Recommendation-only; no prior authorization submission, appeal filing, coverage guarantee, or medical-necessity determination.
• Requires policy source ownership, versioning, reviewer signoff, and payer-action hard stops.

recommendation-search

• Improves reuse of validated patterns while keeping bad or unreviewed recommendations from being retrieved as precedent.
• No autonomous recommendation release; retrieved patterns are examples for human-reviewed synthetic workflow planning only.
• Requires decision-memory retention policy, reviewer outcome labels, and explicit exclusion of hidden chain-of-thought.

I need an appointment or follow-up.

• Blocked: autonomous appointment confirmation, emergency triage replacement, patient outreach without consent
• prepare ranked options and missing-information packet before human action

I need help getting into the right care path.

• Blocked: clinical triage, patient-specific care instruction, live record mutation
• front-load structured intake gaps and owner routing

My referral needs to move forward.

• Blocked: autonomous referral acceptance, patient outreach, clinical urgency determination
• surface delay reasons and next-owner queue

My ordered service needs authorization.

• Blocked: payer submission, coverage guarantee, medical-necessity determination
• reduce packet-prep time while preserving manual verification

My bill, claim, or denial needs resolution.

• Blocked: final coding, billing submission, appeal filing, reimbursement guarantee
• prioritize workqueues and evidence gaps

I need education, reminders, or support.

• Blocked: autonomous patient messaging, medical advice, medication change instruction
• prepare reviewed drafts immediately after approved trigger

I need help navigating the system.

• Blocked: benefit guarantee, clinical advice, identity-dependent action
• route to the right owner with context and boundaries

Real-world performance slices prepared for synthetic evaluation first

• MedLog-style fields: usage_event_id, workflow_trace_id, input_hash, input_classification, model_provider_and_version, output_hash, action_taken, user_feedback, downstream_outcome, fairness_slice, context_effects, clinician_behavior_change
• MLflow-style registries: prompt-registry-production-aliases, trace-registry-with-sme-ground-truth, custom-judge-and-deterministic-scorer-registry, rag-evaluation-registry, model-comparison-registry
• Inference backlog: speculative decoding, block drafting, batching, semantic caching, vLLM/TensorRT-style runtime readiness

Prescription engagement workflow remains review-gated

• Medication education packet draft with source references and plain-language questions for clinician/pharmacist review.
• Reviewable barrier checklist for cost, access, pharmacy pickup, language, and understanding.
• Draft education and follow-up plan for review, with uncertainty and escalation language.
• Task for manual verification when prescription receipt, pickup, or understanding is uncertain.