Storyboard · Healthcare on AWS

Open for the reviewer

Audience level — confirm blended 200/300 or leadership-only (collapses ~40 300 slides).
Per-item rhythm — seven slides each is the proposal; U6 and F0b are lighter and could drop to five.
S66 placeholder — keep as a reserved slide or remove until the portal skeleton exists.
.pptx — HTML deck first; say if a native PowerPoint file is required for a specific audience.
Diagram style — AWS Architecture Icons in the Guidance numbered-step style is the proposal; confirm before D-08 is drawn (it sets the template for the other 34).
AGENT-PLAN — add this storyboard as item 2.0 (before the portal contract) or as a sibling item; it satisfies the docs/architecture/ content-contract line for every item.

Act 0, 5 slides

Frame

S00200

text slide

Title

on screen

title card; one-line subtitle "A HIPAA-grade healthcare platform on AWS, taught through eight working use cases"; date; artifact link.

say

"This is a teaching deck. Every slide is a claim with a source you can click. When we say 'AWS says', the link is on the slide."

sources—

S01200

text slide

Who this is for and how to read a slide

on screen

annotated example slide: citation badge [C35 · High], pattern tag, level tag, lead decision marker, "our pattern — no AWS guidance" marker.

say

"Three kinds of markers. A citation badge means the primary AWS source says this and we re-fetched it on 2026-09-04. A lead-decision marker means the evidence gave us a trade-off and we chose. An 'our pattern' marker means AWS publishes nothing — we are telling you that plainly rather than hiding it."

sourcesdossier header (74 of 78 validated)

S02200

D-01map

The platform in one picture

on screen

D-01 platform map — patients and clinicians on web/iPhone/Android; the eight items as tiles on top of the F0 baseline; arrows showing data flowing U1→U2→U5 and models flowing back.

say

"One platform, eight demonstrations. A patient portal with native apps, devices streaming through phones, visits becoming notes, documents becoming packets, a governed lake, models trained on the platform's own data flowing back into the apps — all on one HIPAA baseline."

sourcesbrief / dossier §1

S03200

D-02pattern

Three tiers per use case

on screen

D-02 three-tier ladder: Tier 1 Core → Tier 2 Embedded AI (Bedrock + AgentCore) → Tier 3 Learned models (SageMaker AI). Same ladder repeated for each item as a mini-icon.

say

"Every use case climbs the same ladder. Tier 1 is the working system. Tier 2 embeds an assistant into that workflow. Tier 3 trains models on the data the system produced. The ladder is why the services are taught through the use case, not as a tour."

patternTiered capability ladder

sourcescatalog-v1 design rules 1–2

S04300

D-03pattern

The stack spine

on screen

D-03 spine diagram: CDK (TypeScript) · FastAPI on ECS Fargate Express Mode · Aurora PostgreSQL Serverless v2 · DynamoDB · Cognito · CloudFront · Bedrock + AgentCore · SageMaker AI; each service a link to its docs page.

say

"One spine everywhere, so a reader who learns U1 already knows 70% of U3. The library records why this spine was chosen for a 100K-user envelope."

patternShared platform spine

sourcescatalog-v1 rule 5library high-scale-app-tier-defaults

Act 1, 6 slides

Ground rules

S05300

text slide

Nothing retired in a new design

on screen

two-column "use / do not use" board: Pinpoint→End User Messaging, Bedrock Agents classic→AgentCore, Kendra→Managed KB, QuickSight→Amazon Quick, Kinesis Data Analytics→Managed Flink, Clarify→SHAP+MLflow, Edge Manager→ONNX+Greengrass v2, Amplify Gen 1→Gen 2, classic Mobile SDKs→Amplify Swift/Android + IoT Device SDK; each with its effective date.

say

"The HIPAA eligibility list is not a lifecycle signal — Forecast, Kendra and Q Business are still listed and still closed to new customers. So every design is checked against the 2026-06 service-availability wave before it is drawn."

patternLifecycle gate

sourcesC12 C20 C33 C59 C15 C17 C77

S06300

D-04pattern

Rule: PHI at the agent tool boundary

on screen

D-04 two-panel diagram. Left: model-call Guardrail sees prompt and completion text only — toolUse.input, toolResult, toolSpec pass untouched. Right: AgentCore Gateway with Cedar Policy on which tools may be called, REQUEST/RESPONSE Lambda interceptors calling ApplyGuardrail, CloudWatch Logs data-protection policy on invocation logs.

say

"This is the single most important correction from the research. A Guardrail on the model call does not see tool inputs or results — AWS says so verbatim. The mechanism that protects PHI in tools is the Gateway: policy on which tools, interceptors on what flows through them, and a logs policy because invocation logs keep the raw prompt. No public healthcare sample of this exists; we build the first."

patternInterceptor at the tool boundary; Policy-as-code (Cedar)

sourcesC74 C75X1G17

S07300

D-05pattern

Rule: model retention modes and PHI

on screen

D-05 retention ladder none < default < aws_review < provider_data_share; which Claude models allow which; the SCP condition key bedrock:DataRetentionMode; arrow showing retained data landing in the destination region.

say

"PHI prompts go to a model that can run with retention none. Fable 5.1 requires aws_review — up to 30 days inside AWS with possible human review — so it is used only on de-identified inputs, and the decision is logged. An SCP pinning none on PHI accounts makes Fable unavailable there by construction. Use a US-only inference profile because retained data lives where the inference ran."

patternRetention gate (preventive control)

sourcesC73 C72

S08300

text slide

Rule: HIPAA-eligible by name vs by feature

on screen

table — named on the Eligible Services Reference (Last Updated 2026-09-03) vs resting on the "GA features of listed services" clause: S3 Tables, Timestream for InfluxDB, Valkey engine, Bedrock Knowledge Bases / Guardrails / Data Automation. Feature-rule sentence quoted verbatim with link.

say

"Every service we name is on the list — except these six, which rest on one sentence. We say so on every chapter that uses them, and until the account team confirms, PHI landing in S3 Tables or InfluxDB is device-keyed and de-identified."

patternEligibility-by-feature disclosure

sourcesC11 C28 C55G2G3

S09300

D-06map

Rule: region and account

on screen

D-06 region map: us-east-1 primary (HealthScribe is here only); us-west-2 secondary for everything else; Device Farm us-west-2 only and never PHI. Call-out: "AWS (advanced)" account + BAA in AWS Artifact before any PHI step.

say

"HealthScribe alone pins us to N. Virginia. And the empty-account runbook cannot start from the new sign-up experience — AWS's own compliance pages say it is not designed for regulated workloads. Advanced account, BAA accepted, then the first PHI byte."

sourcesC35 C78 C25 C7 C8

S10300

text slide

Rule: reuse before rebuild, and label it

on screen

licence census strip: MIT-0 (default) · Apache-2.0 · MIT · CC-BY-SA-4.0 = link only, never fork · none = do not use. Community-maintained repos carry a "community" label.

say

"Fork only permissive licences. The Security Reference Architecture examples are CC-BY-SA — we link them, we never copy. Anything with a retired dependency is on the do-not-link list you will see in the compliance chapter."

patternFork / link / build triage

sourcesC79 C10catalog-v1 rule 3dossier §7

Act 2, 56 slides

The eight items

HIPAA-ready landing zone

Build / LinkS11–S17, 7 slides

S11Story200

D-07journey

Story: from a bare account to audit-ready in one deploy

on screen

D-07 before/after: empty "AWS (advanced)" account → BAA → cdk deploy → account passing the Config HIPAA pack and Security Hub CSPM standards, Macie PHI discovery on, evidence bundle in S3.

say

"The story is deliberately short: one deploy, and an account you could hand to an auditor. Every other item sits on this."

sourcescatalog-v1 §F0 story

S12Exists200

text slide

What exists today

on screen

two cards. LZA + LZA-for-Healthcare: organisation-only ("requires these three accounts at minimum"), ~US$430/month idle, config v1.9.0-e. Single-account HIPAA baseline: nothing public exists (closest is the startup security-baseline pack, 16 controls).

say

"AWS's only official landing-zone artefact needs an Organization and three accounts. No public single-account HIPAA baseline exists anywhere. That absence is why F0 splits: F0a we build, F0b we link."

sourcesC1 C2 C3

S13Tier 1300

D-08architecture

F0a Tier 1 architecture

on screen

D-08 numbered-step architecture: (1) KMS CMKs (2) CloudTrail account trail (3) Config + "Operational Best Practices for HIPAA Security" pack (4) Security Hub CSPM: FSBP + NIST 800-53 r5 + AI Security Best Practices (5) GuardDuty (6) Macie with the 12 PHI identifiers selected explicitly (7) VPC, private subnets, endpoints (8) WAF + Shield (9) Secrets Manager (10) IAM Identity Center single-account (11) CDK Pipelines (12) evidence export to Object-Locked S3.

say

"Twelve steps, one CDK app. Notice step 6: Macie's twelve PHI identifiers are in no default set — accounts that enabled automated discovery before 2023-08-02 behave differently — so the construct selects them explicitly."

sourcesC4 C5 C6 C76

S14Why300

text slide

Why it is built this way

on screen

decision cards: Config HIPAA pack (130-rule cap, last commit 2025-01-07, "not designed to fully ensure compliance"); Security Hub CSPM not "AWS Security Hub" (only CSPM is on the HIPAA list by name); no HIPAA-named Security Hub standard, so FSBP + NIST 800-53 r5 is the practical pairing; evidence export because auditors read S3, not consoles.

say

"Preventive controls are KMS, SCP and VPC. Detective controls are Config, Security Hub and GuardDuty. Evidence is the third leg — findings exported to a bucket with Object Lock that an auditor can read without console access."

patternPreventive + detective + evidence; Guardrails-as-code

sourcesC4 C5 C76

S15300

D-09architecture

F0b: the organisation layer (link, not build)

on screen

D-09 LZA-for-Healthcare on Control Tower: management / audit / log-archive accounts, org-scope BAA from the management account, org trail, SCPs including the DataRetentionMode SCP from S07.

say

"When you have an Organization, AWS already demonstrates this — Landing Zone Accelerator with the healthcare config. Our chapter explains why it is not the first deploy, and contributes one thing: the retention-mode SCP snippet."

sourcesC1 C3 C7 C73G7

S16Compliance300

text slide

The compliance angle

on screen

checklist: BAA scope (account vs organisation) · "AWS (advanced)" prerequisite quoted · Healthcare Industry Lens (2022-11-17, single revision, predates Bedrock) as framing · archived "Architecting for HIPAA" whitepaper points to the eligible-services page.

say

"The Healthcare Lens is still the only official healthcare lens — usable for the shared-responsibility framing, but it predates Bedrock, HealthScribe and AgentCore, so we say when we go beyond it."

sourcesC7 C8 C9

S17Verdict300

text slide

Fork / link / build

on screen

verdict card. F0a: BUILD (no fork base; borrow CDK shape from deploy-conformance-pack-for-aws-startup-security-baseline, MIT-0). F0b: LINK (LZA Solution v1.16.2, LZA-for-Healthcare MIT-0, CfCT Apache-2.0; SRA examples CC-BY-SA link-only). Demonstration value: F0a High, F0b Low.

say

"F0a is a genuine contribution — nothing public exists and everything depends on it. It is also small, which is why it is built first."

sourcesC2 C10 C1 C3

Patient portal + companion mobile app

BuildS18–S24, 7 slides

S18Story200

D-10journey

Story: one patient journey on three clients

on screen

D-10 journey strip: sign up → passkey or MFA → appointments and lab results → message the care team → upload a document → reminder arrives; the same strip on web, iPhone, Android; every PHI access logged.

say

"This is the reference application. Everything else inherits its decisions: how a patient signs in, how the API is shaped, how notifications travel, where the clinical record lives."

sourcescatalog-v1 §U1 story

S19Exists200

text slide

What exists today

on screen

absence card: no AWS-official portal + native mobile + Cognito + HealthLake example on any stack (three independent searches). Component cards: sample-intelligent-security-for-healthcare-apis (API boundary, Cognito MFA, Comprehend Medical redaction before audit logging); aws-healthlake-smart-on-fhir (FHIR wiring, no UI, external IdP).

say

"Highest-value gap in the catalog. Two samples give us the API boundary and the FHIR wiring; everything the user touches is new."

sourcesC13 C14

S20Tier 1300

D-11architecture

Tier 1 architecture

on screen

D-11 numbered steps: (1) React 19 + Vite on S3/CloudFront (2) Swift and Kotlin apps on Amplify Gen 2 (3) Cognito user pool, Essentials plan, Managed Login (4) API Gateway HTTP API → FastAPI on Fargate Express Mode (5) AppSync Events for real-time (6) Aurora Serverless v2 app state (7) S3 + KMS documents (8) HealthLake FHIR R4 record, stack toggle (9) EventBridge → SNS mobile push / End User Messaging SMS / SES email.

say

"Nine steps. HealthLake is the dominant running cost at about US$197 a month per data store, so it is a stack toggle you can turn off between demos."

sourcesC18 C19 C24 C21

S21Why300

D-12decision

Why it is built this way

on screen

D-12 decision board with four cards. (a) Native Swift + Kotlin on Amplify Gen 2 lead decision — Flutter lacks passkeys and AppSync Events (issues open since 2025-03/04), Gen 1 EOL 2027-05-01, the U2 device path is native anyway. (b) Passkey or password+MFA, never OTP-first — "OTP flows aren't compatible with required MFA"; passkeys need MULTI_FACTOR_WITH_USER_VERIFICATION and the .well-known association file. (c) API Gateway to proxy an existing backend + AppSync Events for real-time — no AWS doc prefers one over the other; this is the documented split. (d) SNS mobile push as default lead decision — End User Messaging Push still keys on a Pinpoint application resource whose survival past 2026-10-30 is unconfirmed.

say

"Four decisions the rest of the catalog inherits. Each one has an ADR in the repo; each one is a trade-off the sources describe but could not settle for us."

patternBackend-for-frontend proxy; Passkey-first authentication; Event fan-out

sourcesC15 C16 C19 C18 C20 C21X2G18

S22T2 + T3300

D-13architecture

Tier 2 and Tier 3

on screen

D-13 Tier 2: in-app assistant — Bedrock + Managed Knowledge Base over patient-education content + AgentCore Gateway tools that read this patient's FHIR data, wrapped in the S06 interceptor pattern; AgentCore Memory KMS-encrypted; Macie/Comprehend Medical scan of KB content before ingestion. Tier 3: no-show model (XGBoost, Feature Store) → batch score to DynamoDB → reminder.

say

"Tier 2 is 'what does this result mean?' answered from the patient's own record with PHI protected at the tool boundary. Tier 3 is the first model on the SageMaker Pipelines spine — a no-show predictor whose only visible output is a reminder."

patternTools not embeddings; Pipelines spine

sourcesC22 C73 C74 C75 C23 C59

S23Compliance300

D-14sequence

The compliance angle

on screen

D-14 "notification without payload" sequence: result ready → EventBridge → SNS → APNs/FCM carries only "you have a new result" → app fetches over the authenticated API. Side notes: APNs/FCM are outside the BAA; every FHIR read is logged; Comprehend Medical redaction before audit logging (from the forked sample).

say

"Push services are not under the BAA, so the notification is a doorbell, not a letter. The app fetches the PHI itself, authenticated, logged."

patternClaim check (notification without payload)

sourcesC21 C14

S24Verdict300

text slide

Fork / link / build

on screen

verdict card. Tier 1 BUILD — fork sample-intelligent-security-for-healthcare-apis (MIT-0, CloudFormation → CDK TS conversion) and aws-healthlake-smart-on-fhir (MIT-0, CDK TS). Tier 2 FORK scaffolding, BUILD assistant — HCLS Agents Toolkit, sample-healthcare-agent-with-agentcore-on-aws, public-sector blueprint blog. Tier 3 BUILD on amazon-sagemaker-pipeline-deploy-manage-100x-models-python-cdk. Value: Highest / High / Medium.

say

"New work is everything user-facing on three clients, the passkey flow, the FastAPI service, HealthLake on Cognito rather than ORY, and the fan-out."

sourcesC14 C22 C23

Remote patient monitoring from connected devices

BuildS25–S31, 7 slides

S25Story200

D-15journey

Story: vitals through the phone, trends for the care team

on screen

D-15 journey: cuff / meter / wearable → patient's phone (HealthKit, Health Connect, BLE) → cloud → care-team trend view and anomaly alert → patient sees own data and asks the assistant → models trained on fleet history score new readings in the cloud and at the edge.

say

"The device-data showcase. Everything here is new work — there is no end-to-end RPM example on AWS, and no AWS guidance for HealthKit or Health Connect at all."

sourcescatalog-v1 §U2 storyC26

S26Exists200

text slide

What exists today

on screen

component cards with dates: Greengrass foundations Guidance (CDK TS, 2026-07-23) · EC2 device farm for a simulated fleet (2026-08-30) · AppSync IoT real-time dashboard (Amplify Gen 2, 2026-09-03) · BLE Swift sample (CDK v2 current, iOS layer on an EOS SDK — CDK only) · Flink keyed Random Cut Forest (no IaC) · DeepAR MLOps pipeline. Absence card: no IoT Core → Timestream for InfluxDB sample with IaC.

say

"Pieces exist; the composition does not. And the two stores we want — InfluxDB hot, S3 Tables cold — have no healthcare sample and are eligible only by the feature rule."

sourcesC27 C28 C31X6X7

S27Tier 1300

D-16architecture

Tier 1 architecture

on screen

D-16 numbered steps: (1) phone authenticates with Cognito identity pool + IAM role + IoT policy attached to the identity, MQTT 5 over WebSocket/SigV4 (2) IoT Core, X.509 per gateway device, Device Shadows, Fleet Provisioning (3) IoT Rules → Kinesis Data Streams (4) Managed Service for Apache Flink (5) Timestream for InfluxDB hot store, device-keyed, de-identified (6) Firehose → S3 Tables (Iceberg) cold store → Athena (7) Greengrass v2 gateway with offline buffering (8) EventBridge alerts → SNS push / End User Messaging SMS (9) portal device UI + real-time view.

say

"Hot store for the last days, cold store for the fleet's history, one Flink job in between doing thresholds now and anomaly scoring later. The phone is a first-class device with its own IoT policy."

patternHot/cold store; Stream processing with keyed state; Device shadow

sourcesC29 C27 C28 C34 C11

S28Why300

D-17pattern

Why it is built this way: the phone is the gateway

on screen

D-17 phone-as-gateway diagram: HealthKit (on-device only — Apple 5.1.3 forbids iCloud storage of health data) and Health Connect (background reads + sync) read by the app; BLE devices bridged by the phone; single MQTT path to IoT Core. Big marker: "AWS publishes no guidance for this — it is our pattern." iOS: AWS IoT Device SDK for Swift (GA 2026-06-24); Android: IoT Device SDK for Java v2; Amplify has no PubSub category.

say

"Apple will not let health data touch iCloud, and Android's Health Connect is an on-device store too. So the phone reads locally and forwards over MQTT with its Cognito identity. AWS documents each half; nobody documents the join. We say that on the slide."

patternEdge gateway (phone as aggregator)

sourcesC29 C30 C17G11G19

S29T2 + T3300

D-18architecture

Tier 2 and Tier 3

on screen

D-18 Tier 2: telemetry exposed as typed MCP tools via AgentCore Gateway (not embeddings); Cedar restricts each patient principal to their own device tools; interceptors mask PHI on the way through. Tier 3: Random Cut Forest in Flink state → anomaly alert; DeepAR deterioration forecast; retraining on drift through Pipelines + Model Registry; edge = quantised ONNX model as a Greengrass v2 component (Edge Manager is gone).

say

"The assistant asks 'how has my BP trended?' through a typed tool, never by embedding raw telemetry. And the first public healthcare example of interceptor-based PHI masking lives here."

patternTools not embeddings; Interceptor at the tool boundary; Edge scoring

sourcesC32 C74 C75 C33 C34 C27

S30Compliance300

text slide

The compliance angle

on screen

table: what lands where — identifiers in Aurora/HealthLake; hot and cold stores hold device-keyed, de-identified rows until G2/G3 are closed; Device Farm (us-west-2) never receives PHI; Greengrass listed on the HIPAA reference without a version qualifier.

say

"Two open gaps — S3 Tables and InfluxDB by name — decide how much PHI the stores may hold. Until the account team answers, the pipeline is designed so it does not matter."

sourcesC11 C28 C55 C25G2G3

S31Verdict300

text slide

Fork / link / build

on screen

verdict card. All tiers BUILD. Fork: Greengrass foundations (MIT-0), EC2 device farm (Apache-2.0), AppSync IoT dashboard (MIT-0, Gen 2), BLE Swift sample CDK only; RCF logic (add IaC), DeepAR pipeline. Do not link: Edge Manager sample, archived Greengrass-v1 forecasting. New: Flink job, InfluxDB + S3 Tables landing with IaC, Kotlin and Swift device clients, alerting, device UI, load generator to the scale target. Value: Highest / High / High.

say

"Highest demonstration value in the catalog because every tier is a gap."

sourcesC26 C27 C28 C29 C30 C31 C32 C33 C34 C79

Ambient clinical documentation

Fork + buildS32–S38, 7 slides

S32Story200

D-19journey

Story: a recorded visit becomes an approved note

on screen

D-19 journey: clinician records on phone or web → SOAP note + ICD-10 suggestions + patient-friendly summary, each sentence with a timestamp citation → clinician reviews and approves → FHIR DocumentReference in HealthLake.

say

"The point of this demo is the human in the loop. The machine drafts with citations; the clinician approves; only then does it become part of the record."

sourcescatalog-v1 §U3 story

S33Exists200

D-20decision

What exists today: two products, one choice

on screen

D-20 option-pair table. A — HealthScribe direct: us-east-1 only, en-US, US$0.10/min, CDK TS sample that deploys empty→running, but the dev guide banner steers to Connect Health. B — Amazon Connect Health Ambient: us-east-1 + us-west-2, US$99/user/month (600 encounters), needs a Connect instance + Connect Health domain, best sample uses Bedrock Agents classic.

say

"AWS gives you two ways to do the same thing, and the newer one is a subscription bolted to a contact-centre product. This is the first real trade-off in the deck."

sourcesC35 C37 C38 C39 C60X3

S34Tier 1300

D-21architecture

Tier 1 architecture (Option A)

on screen

D-21 numbered steps: (1) capture → S3 (2) HealthScribe streaming or batch (3) Comprehend Medical entities + ICD-10-CM / RxNorm / SNOMED linking (4) Step Functions review workflow (5) clinician review UI states (6) HealthLake DocumentReference write-back. Fallback lane: Transcribe Medical + Bedrock for other regions.

say

"Fork the HealthScribe sample for structure and UI; port the coding step from the diagnosis-codes Guidance; build the review workflow and the write-back — the two things no sample has."

patternHuman-in-the-loop approval; Provenance / citation

sourcesC35 C36 C39 C40

S35Why300

text slide

Why it is built this way lead decision

on screen

decision card: Option A chosen because it deploys from an empty account on the catalog's own spine, costs cents per demo, and stays independent of U6; accept the us-east-1 pin; Option B is demonstrated as a product in U6; revisit if HealthScribe gets an end-of-support notice ("steered, not retired"). Composition note (X4): CDK-TS sample for structure + Comprehend Medical slice from the Guidance repo (only that slice is in its code; drop its US$350/month OpenSearch Serverless line for a Managed KB).

say

"We chose the path that teaches the most for the least money, and we wrote down the condition under which we would switch."

patternOption pair with a documented switch condition

sourcesX3X4C40 C78

S36T2 + T3300

text slide

Tier 2 and Tier 3

on screen

Tier 2: Bedrock drafts SOAP note and patient summary with timestamp citations; Guardrail on the model call (text) plus S06 pattern if tools are used; visit transcripts are PHI → ZDR-capable model per S07; the sample's 12-specialist Step Functions orchestration forked. Tier 3: note-quality / coding-suggestion classifier; evaluation harness from the LLM-as-judge notebooks — "when does a trained model beat a prompt?"

say

"Tier 3 here is honest about a question people avoid: is the classifier better than the prompt? The harness answers it with numbers."

patternEvaluation harness

sourcesC39 C73 C74 C42

S37Compliance300

text slide

The compliance angle

on screen

cards: HealthScribe is on the list as "AWS Transcribe [Includes Healthscribe]"; Comprehend Medical DetectPHI returns entities with offsets, does not redact, and per AWS "does not meet the requirements for de-identification" — automation assists, Safe Harbor is a human process; retention mode for transcripts; region pin.

say

"The de-identification sentence is quoted because people assume the opposite. Comprehend Medical finds PHI; it does not make a document de-identified."

sourcesC11 C43 C73 C35

S38Verdict300

text slide

Fork / link / build

on screen

verdict card. Tier 1–2 FORK + BUILD — sample-healthscribe-bedrock-clinical-analysis (MIT-0, CDK TS, 2026-04-07); port from guidance-for-identifying-diagnosis-codes-from-clinical-notes-on-aws (MIT-0, CDK Python); UI ideas only from aws-healthscribe-demo (Gen 1). Tier 3 BUILD. Open: Comprehend Medical unit prices (G6). Value: Medium across tiers.

say

"Fork bases exist; the composition is the value."

sourcesC39 C40 C41 C42G6

Intelligent document intake

Link + forkS39–S45, 7 slides

S39Story200

D-22journey

Story: from fax to prior-auth packet

on screen

D-22 journey: PDFs and faxes arrive → extracted, classified, routed → PHI detected and masked for logging → human resolves exceptions → agent assembles a prior-auth packet.

say

"The best-covered item in the catalog — which is itself the lesson: know when AWS has already built it."

sourcescatalog-v1 §U4 storyC51

S40Exists200

text slide

What exists today

on screen

cards: GenAI IDP Accelerator (pushed 2026-09-04, MIT-0, 302 stars; one-click CloudFormation in us-east-1/us-west-2/eu-central-1; Pattern 1 = Bedrock Data Automation, Pattern 2 = Textract → Bedrock; built-in human review, MLflow evaluation, MCP via AgentCore Gateway) · Guidance for IDP on AWS page (AgentCore Runtime/Identity/Gateway + Strands + Textract; its repo fork-blocked until the LICENSE text is confirmed) · sample-healthcare-agents (prior auth, eligibility, coding, claims, appeals; payor-policy PDF tools; no patient-document intake) · two official IDP workshops.

say

"Three current, MIT-0, AgentCore-era artefacts. Our new work shrinks to the healthcare layer."

sourcesC44 C45 C46 C47X8

S41Tier 1300

D-23architecture

Tier 1 architecture

on screen

D-23 numbered steps: (1) S3 intake with Macie PHI identifiers selected explicitly (2) Bedrock Data Automation (Pattern 1) or Textract (Pattern 2, on the HIPAA list by name) (3) Comprehend Medical (4) de-identification stage (aws-ai-phi-deidentification) (5) Step Functions + SQS (6) confidence-gated review queue, web + mobile (7) Aurora / HealthLake landing.

say

"Two patterns side by side so the chapter can show why BDA is the 2026 default and why Textract stays available: one is on the list by name, the other by the feature rule."

patternConfidence-gated review queue; De-identification stage

sourcesC47 C6 C49

S42Why300

text slide

Why it is built this way

on screen

decision cards: link the Accelerator rather than fork (it is the deployable; our value is the healthcare layer) · embed the Guidance page as the architecture explainer · do not link the BDA sample that uses A2I + Ground Truth (maintenance) · state that automated de-identification assists but does not satisfy Safe Harbor.

say

"The pattern here is composition discipline: link what is maintained, fork what you must change, build only the layer that is yours."

patternLink + fork composition

sourcesC44 C45 C48 C43

S43T2 + T3300

text slide

Tier 2 and Tier 3

on screen

Tier 2: Bedrock structured extraction with the Accelerator's review queue; AgentCore agent assembling the prior-auth packet from sample-healthcare-agents (Strands, HealthLake FHIR R4, B2B Data Interchange 837P, CDS Hooks), wired to the Accelerator's output, under the S06/S07 rules. Tier 3: document-type classifier and denial-risk model on SageMaker from intake history — the Accelerator classifies with Bedrock, so this is the "trained model vs prompt" comparison again.

say

"The Tier 2 agent already exists for payors; we give it patient documents to read."

patternTools not embeddings; Evaluation harness

sourcesC46 C50

S44Compliance300

text slide

The compliance angle

on screen

cards: Macie on the intake bucket (explicit identifiers) · BDA eligibility by feature rule stated on the chapter · Comprehend Medical de-id caveat · Guidance usage disclaimer verbatim ("You should not use this AWS Content in your production accounts") on every forked chapter · Clario blog numbers for Bedrock PHI detection with human review.

say

"Every forked chapter carries AWS's own disclaimer. It is not decoration; it is the licence's condition."

sourcesC6 C47 C43 C68 C49

S45Verdict300

text slide

Fork / link / build

on screen

verdict card. Tier 1 LINK Accelerator; Tier 2 FORK sample-healthcare-agents (MIT-0, CDK TS cdk deploy --all); add aws-ai-phi-deidentification (MIT-0, CDK); Guidance repo fork blocked (licence). Tier 3 BUILD. Value: Low / Low-Medium / Medium.

say

"Two chapters for little code, and it exercises the portal's link and fork paths early — which is why it is fifth in the build order."

sourcesC44 C45 C46 C47 C48 C49 C50 C79

Population health analytics and risk models

Fork + buildS46–S52, 7 slides

S46Story200

D-24journey

Story: a governed lake you can ask questions of

on screen

D-24 journey: data from U1–U4 lands in a governed lake → analysts see de-identified Quick dashboards → ask the lake a question in plain language → risk models score patients → flags appear back in U1.

say

"This is the data platform. It needs U1 to U4 to be interesting, which is why it comes late."

sourcescatalog-v1 §U5 story

S47Exists200

text slide

What exists today

on screen

cards: Multi-Modal Data Analysis Guidance (2024-12-17, MIT-0; notebooks; HealthOmics/HealthLake/HealthImaging → Lake Formation → Athena → QuickSight; Feature Store → AutoGluon → endpoint; no S3 Tables, no de-id, no cost table) · Patient Entity Resolution Guidance (2025-09-11, MIT-0) = EMPI · healthlake-workshop archived 2026-02 · readmission samples all 2021–2022 · extract-medical-insights-… text-to-SQL snippet, no IaC.

say

"Good bones, old naming. Every artefact still says QuickSight."

sourcesC53 C54 C56 C57 C58

S48Tier 1300

D-25architecture

Tier 1 architecture

on screen

D-25 numbered steps: (1) HealthLake export + Firehose/Iceberg from U2 (2) S3 Tables (Iceberg) with the feature-rule caveat (3) Glue Data Catalog (4) Athena (5) Lake Formation fine-grained access (6) AWS Entity Resolution as EMPI (7) de-identification pipeline: Comprehend Medical DetectPHI + Macie + de-id sample, human review documented (8) Amazon Quick dashboards.

say

"Governance is a service here, not a policy document: Lake Formation decides who sees which column."

patternGoverned lakehouse; EMPI via entity resolution

sourcesC53 C54 C55 C56 C43 C49

S49Why300

text slide

Why it is built this way

on screen

decision cards: S3 Tables as the cold store shared with U2 (one lake, two producers) · Quick rename pass required across every forked artefact · de-identification is a process with human review, not a function · CDK TS conversion and a cost table because the Guidance has neither.

say

"The Guidance gives us the data model and the notebooks. We give it infrastructure as code, a cost table, and the parts that make it HIPAA-shaped."

patternSingle lake, many producers

sourcesC53 C56 C43

S50T2 + T3300

D-26architecture

Tier 2 and Tier 3

on screen

D-26 Tier 2: natural-language analytics — Quick, or Bedrock + Athena tools via AgentCore Gateway under S06; Managed KB over care guidelines. Tier 3: readmission / chronic-condition risk (Feature Store, Pipelines, Model Registry) → batch scores to DynamoDB → flags in U1; bias and explainability = standardised bias metrics + SHAP logged to SageMaker AI MLflow (Clarify successor); monitoring from sample-aiops-on-amazon-sagemakerai.

say

"Clarify is closed to new customers, and AWS names its replacement: open metrics, SHAP, MLflow. That is what the pipeline logs, and the flag in the patient app is traceable to a registered model."

patternPipelines spine; Model registry; Explainability in the pipeline

sourcesC58 C59 C57X5

S51Compliance300

text slide

The compliance angle

on screen

cards: S3 Tables by feature rule (G2) · Lake Formation column-level access · de-identified rows only until G2/G3 close · Quick on the HIPAA list by name · Safe Harbor is a human-reviewed process.

say

"Same two open gaps as U2, same design answer: de-identified until told otherwise."

sourcesC55 C56 C43G2

S52Verdict300

text slide

Fork / link / build

on screen

verdict card. Tier 1 FORK + BUILD Multi-Modal Guidance + Entity Resolution Guidance; build S3 Tables landing, Quick, de-id, CDK TS, cost table. Tier 2 BUILD. Tier 3 FORK content, BUILD infrastructure [lead decision X5] on the 100x-models CDK spine. Value: Medium / Medium-High / Medium.

say

"Tier 2 is Medium-High because nobody has shown natural-language analytics over a HIPAA lake."

sourcesC53 C54 C57 C58 C59X5

Patient engagement contact centre

LinkS53–S58, 6 slides

S53Story200

D-27journey

Story: the patient calls

on screen

D-27 journey: patient calls or chats → Connect Health verifies identity → scheduling and FAQs → Bedrock-backed agent answers policy questions → escalation to a human carries context → the same Connect Customer campaigns send U1's reminders.

say

"This chapter demonstrates an AWS product rather than building one. That is a deliberate lesson: when AWS makes something a product feature, building it yourself demonstrates the wrong thing."

sourcescatalog-v1 §U6 storyC60

S54Exists200

text slide

What exists today

on screen

cards: Amazon Connect Health GA 2026-03-05 (patient verification GA US$0.15/action; ambient documentation GA US$99/user/month; appointment management preview; medical coding gated preview) · sample-healthcare-realtime-eligibility (MIT-0) · Nova Sonic call-centre sample (MIT, no Connect) · samples on Bedrock Agents classic — do not link · no healthcare Connect workshop or Guidance.

say

"Verification and scheduling are now first-party features. Lex plus Lambda identity code would be teaching 2023."

sourcesC37 C60 C61 C62 C63 C64

S55Tier 1300

D-28architecture

Tier 1 as a product walkthrough

on screen

D-28 Connect Customer instance + Connect Health domain; Lex V2 where needed; web chat widget + in-app chat; End User Messaging SMS; SES email; eligibility Lambda; Connect Health Ambient shown here as U3 Option B.

say

"Console-driven by design — a Connect instance is not empty→running by CDK alone, so the chapter shows it running and explains what it would cost."

sourcesC60 C61 C37

S56T2 + T3300

text slide

Tier 2 and Tier 3

on screen

Tier 2: Connect Health agentic features + KB over policies; optional fork of the Nova Sonic AgentCore sample if a deployable repo is wanted; serverless-patient-engagement-stack reusable only after Pinpoint is replaced. Tier 3: demand forecasting / capacity / scheduling is a Connect product feature — link; contact-reason classification could be a SageMaker build.

say

"Tier 3 here is a link, and that is the honest verdict."

sourcesC63 C65 C20 C64

S57Compliance300

text slide

The compliance angle and the U1 connection

on screen

cards: Amazon Connect and Lex on the HIPAA list · Connect Health HIPAA-eligible in both regions · engagement successor to Pinpoint is Connect Customer campaigns + Customer Profiles, so U1 reminders and U6 share a service.

say

"One consequence of the Pinpoint retirement: the contact centre and the patient app share their engagement plumbing."

patternShared engagement channel

sourcesC61 C37 C21 C65

S58Verdict300

text slide

Fork / link / build

on screen

verdict card. Tier 1–2 LINK the product; optional FORK Nova Sonic sample (MIT, CDK Python). Tier 3 LINK. Value: Low.

say

"Last in the build order, mostly prose — and still a chapter worth having."

sourcesC60 C61 C62 C63 C64 C65

The education portal

BuildS59–S66, 8 slides

S59Story200

D-29map

Story: the frame every other item is seen through

on screen

D-29 portal shape: one site; a chapter per item; each chapter links the live demo, the repo, the CDK stack list, the architecture diagram, and a "deploy this yourself" runbook; forked chapters carry the Guidance disclaimer.

say

"Nothing like this exists publicly for this purpose. But AWS's content surfaces have stable contracts, and we copy them rather than invent."

sourcesC66 C67 C68 C69 C70 C71

S60Exists300

text slide

What AWS's three surfaces teach us

on screen

three-column anatomy: Workshop Studio (contentspec.yaml v2.0, accountSources, templates under static/) · Solutions Library Guidance page (overview → diagram + numbered steps → sample code → six pillars → related → disclaimer; cost and deploy steps live in the repo README) · aws-samples README (MIT-0; Overview → Cost → Prerequisites → Deployment → Validation → Running → Next Steps → Cleanup; _pattern-model/example-pattern.json as the smallest docs-from-metadata contract).

say

"Three copyable page anatomies. Our chapters use the Guidance anatomy, the Solution header for maintenance signals, and the serverless-patterns metadata file to drive tables."

sourcesC66 C68 C69

S61300

D-30pattern

The content contract

on screen

D-30 repo-to-portal diagram: every use-case repo ships README.md (Guidance contract) · catalog-item.json (id, tiers, services with HIPAA entry name and eligibility basis, regions, fork provenance, retired-service check date) · docs/decisions/ (MADR 4) · docs/architecture/ (diagram source + PNG, numbered steps) · contentspec.yaml; CI emits the CDK stack list, cost rows and retired-service audit; the portal aggregates at build time and holds navigation, not prose.

say

"Repo-owned docs, portal-aggregated — the Backstage TechDocs model. The portal never drifts because it never holds prose of its own."

patternDocs-as-code; Content contract; Generated, never hand-written

sourcesC69 C70catalog-v1 §P0

S62Why300

text slide

Why it is built this way lead decision

on screen

decision card: Docusaurus site generated from the use-case repos (precedent eks-workshop-v2, Apache-2.0), on the same CloudFront stack; Backstage TechDocs / Antora as alternatives if the catalog outgrows one site; link-outs to Workshop Studio provisional until module lists are read in a browser (G1).

say

"AWS's own largest open workshop is not on Workshop Studio — it is Docusaurus in a repo. We follow that precedent."

sourcesC70 C67G1

S63300

text slide

ADRs: how decisions stay honest

on screen

MADR 4 file shape (docs/decisions/NNNN-title.md, front matter status/date/decision-makers; Context → Drivers → Options → Outcome → Consequences → Confirmation); list of the ADRs this deck implies: mobile stack, push channel, U3 option, hot-store PHI handling, model/retention choice, single-account landing zone, region, scale target.

say

"Every lead decision marker in this deck is an ADR file. Accepted ADRs are immutable; when we change our mind, we write a new one that supersedes."

patternArchitecture Decision Records

sourcesC70AWS Prescriptive Guidance on ADRs (angle 4 E1)

S64300

text slide

Link-out policy and labelling

on screen

policy cards: AWS workshops and Guidance linked only after their module list has been read · community repos labelled "community" · forked chapters carry the usage disclaimer · "runbook-only" label for snippets without IaC.

say

"Labels are how a reader knows what kind of trust to extend to a link."

sourcesC67 C68D5 (angle 4)brief OQ1

S65Verdict300

text slide

Fork / link / build

on screen

verdict card. BUILD on existing contracts; precedents eks-workshop-v2, HCLS Agents Toolkit site (Astro). Value: High — it is the frame.

say

"Built first, alongside F0a, because retrofitting the contract is the most expensive mistake available."

sourcesC70catalog-v1 build order 1

S66200

reserved

(reserved) Portal walkthrough screenshot

on screen

placeholder for a real screenshot once the skeleton deploys (AGENT-PLAN item 2.2).

say

no narration

sources—

Act 3, 6 slides

Cross-cutting

S67300

D-31pattern

Assistants everywhere: one pattern, four workflows

on screen

D-31 the S06 interceptor pattern instantiated four times: U1 "what does this result mean?" · U2 "how has my BP trended?" · U4 "assemble the prior-auth packet" · U5 "ask the lake a question" — same Gateway, same Cedar, same interceptors, different typed tools.

say

"Build the pattern once in U1 Tier 2; every later assistant is a new set of tools, not a new security design."

patternInterceptor at the tool boundary (reused)

sourcesC74 C75catalog-v1 build order 6

S68300

D-32architecture

Learned models: one spine, five models

on screen

D-32 SageMaker Pipelines + Model Registry + Feature Store + MLflow deployed once; no-show (U1), anomaly + forecast + edge (U2), readmission risk (U5), note quality (U3), document type (U4) flowing through it; scores landing in DynamoDB and surfacing in the apps.

say

"The spine is built once in U1 Tier 3 and reused. Each model is traceable from the flag in the app back to a registered version with its SHAP explanation."

patternPipelines spine; Model registry

sourcesC23 C57 C59catalog-v1 build order 9

S69300

text slide

Scale and operations

on screen

cards: load test to 1.5× peak on the agreed envelope · AppSync Events / WebSocket API Gateway for streaming · ElastiCache for Valkey fan-out (feature-rule eligibility) · RDS Proxy · DynamoDB session state · OpenTelemetry → CloudWatch/X-Ray · AgentCore Observability · blue/green · Alembic-gated migrations · cost dashboard with HealthLake ≈US$197/month as the line to watch (reconcile with the "~$500/month" README figure, G21).

say

"The scale target is one of the four decisions still open; everything on this slide is sized to it."

sourcescatalog-v1 cross-cuttingC24G21library capacity-quotas-and-sizing

S70300

D-33sequence

The compliance-evidence chapter

on screen

D-33 the negative test: a model-call Guardrail missing PHI in toolResult next to the interceptor catching it; the retention-mode policy blocking a non-compliant model; Security Hub CSPM AI Security Best Practices findings; Macie identifier selection; the eligibility-by-feature table; the F-15 closure ("your content is not used to improve the base models"); the "AWS (advanced)" requirement.

say

"This chapter is runnable tests, not prose. The demo that matters most is the one where the wrong control fails on screen."

patternNegative test as evidence

sourcesC72 C73 C74 C75 C76 C6 C8 C11AGENT-PLAN item 7.2

S71200

D-34map

The build order and why

on screen

D-34 dependency graph: P0 contract + F0a → U1 T1 → U2 T1 → U4 T1–2 → U1 T2 + compliance chapter → U3 → U5 T1–2 → Tier 3 across U1/U2/U5 (then U3/U4) → U6 + F0b.

say

"Contract first so nothing is retrofitted; foundation second because everything sits on it; the reference app third because it fixes every inherited decision; the cheap win fourth to exercise link and fork early."

sourcescatalog-v1 §Recommended build order

S72300

text slide

Do not link as-is

on screen

table from dossier §7: repo · retired dependency · what is still usable (e.g. serverless-patient-engagement-stack → Pinpoint; aws-amplify-cdk-iot-ble-swift-app iOS layer → classic SDK; medical-mobile-iot-with-aws → CC-BY-SA and archived).

say

"Every one of these looked usable in a search result. The lifecycle gate is why they are on this slide and not in the architecture."

patternLifecycle gate

sourcesC77 C79

Act 4, 3 slides

Close

S73200

text slide

Four decisions to confirm before we start

on screen

four cards with the v1 default and the alternative: (1) single account for F0 (default: F0a mandatory, F0b optional) (2) us-east-1 primary (default: yes; alternative: Connect Health Ambient for U3 to keep us-west-2 parity) (3) scale target (default: 100K-user / tens-of-thousands-concurrent envelope) (4) two native mobile codebases (default: Swift + Kotlin; alternative: React Native, no Expo Go). Two riders: PHI in InfluxDB / S3 Tables before G2/G3; community repos linkable.

say

"These are yours to confirm. The defaults are already in the catalog; AGENT-PLAN item 1 records the answers as ADRs."

sourcescatalog-v1 §Four ambiguitiesAGENT-PLAN item 1

S74200

text slide

What you can do with this

on screen

three paths: read the catalog (catalog-v1.md) · read the evidence (dossier.md, 79 claims) · follow the plan (AGENT-PLAN.md, 12 items).

say

"Everything on these slides is in the repo, cited, and dated. When a source changes, the claim id tells you which slide to fix."

sources—

S75200

text slide

Sources

on screen

full clickable index grouped as the dossier §9 groups them: compliance · landing zone · Bedrock/AgentCore · health services · mobile/device · SageMaker · repositories · Solutions Library and portal patterns.

say

no narration

sourcesAppendix C

Appendix A — Diagram inventory

Architecture and map diagrams as hand-authored SVG with the official AWS Architecture Icons in the Guidance numbered-step style; sequences as mermaid; journeys as simple strips. All theme-aware. No screenshots exist yet.

id	type	title	used on
D-01	map	Platform in one picture	S02
D-02	pattern	Three-tier ladder	S03
D-03	pattern	Stack spine	S04
D-04	pattern	Guardrail scope vs Gateway interceptor (PHI at the tool boundary)	S06, S67
D-05	pattern	Retention-mode ladder and SCP gate	S07
D-06	map	Region and account constraints	S09
D-07	journey	F0a before/after	S11
D-08	architecture	F0a single-account baseline, 12 numbered steps	S13
D-09	architecture	F0b organisation layer on LZA	S15
D-10	journey	U1 patient journey on three clients	S18
D-11	architecture	U1 Tier 1, 9 numbered steps	S20
D-12	decision	U1 four inherited decisions	S21
D-13	architecture	U1 Tier 2 assistant + Tier 3 no-show model	S22
D-14	sequence	Notification without payload (claim check)	S23
D-15	journey	U2 device-to-care-team journey	S25
D-16	architecture	U2 Tier 1 pipeline, 9 numbered steps	S27
D-17	pattern	Phone as gateway (HealthKit / Health Connect / BLE → MQTT)	S28
D-18	architecture	U2 Tier 2 typed telemetry tools + Tier 3 cloud/edge scoring	S29
D-19	journey	U3 visit-to-note	S32
D-20	decision	HealthScribe direct vs Connect Health Ambient	S33
D-21	architecture	U3 Tier 1 with human-in-the-loop, 6 steps + fallback lane	S34
D-22	journey	U4 fax-to-packet	S39
D-23	architecture	U4 Tier 1 two-pattern intake, 7 steps	S41
D-24	journey	U5 lake-to-flag	S46
D-25	architecture	U5 Tier 1 governed lake, 8 steps	S48
D-26	architecture	U5 Tier 2 NL analytics + Tier 3 risk model with SHAP/MLflow	S50
D-27	journey	U6 the patient calls	S53
D-28	architecture	U6 Connect Customer + Connect Health product view	S55
D-29	map	Portal shape	S59
D-30	pattern	Content contract: repo → CI → portal	S61
D-31	pattern	One interceptor pattern, four assistants	S67
D-32	architecture	One Pipelines spine, five models	S68
D-33	sequence	The negative test: Guardrail misses `toolResult`, interceptor catches it	S70
D-34	map	Build-order dependency graph	S71
D-35	sequence	Passkey sign-in with `MULTI_FACTOR_WITH_USER_VERIFICATION` (optional, S21 backup)	S21

Appendix B — Pattern glossary

Every pattern named on a slide, with the slides that teach it.

pattern	definition	taught on
Tiered capability ladder	Core → Embedded AI → Learned models, same rungs for every use case	S03
Shared platform spine	One compute/data/identity/delivery stack reused by every item	S04
Lifecycle gate	Check every service against retirement/maintenance status before drawing it	S05, S72
Interceptor at the tool boundary	Protect PHI in tool I/O with Gateway REQUEST/RESPONSE interceptors, not a model-call Guardrail	S06, S29, S67
Policy-as-code (Cedar)	Declare which principal may call which tool	S06
Retention gate	SCP/IAM condition on `DataRetentionMode` decides which models PHI may reach	S07
Eligibility-by-feature disclosure	Name the services that rest on the GA-features clause and design so it does not matter	S08, S30, S51
Fork / link / build triage	Reuse by licence and lifecycle; label provenance	S10, every 7th slide
Preventive + detective + evidence	KMS/SCP/VPC · Config/Security Hub/GuardDuty · exported findings	S14
Backend-for-frontend proxy	API Gateway in front of an existing container backend; AppSync Events for real-time	S21
Passkey-first authentication	Passkey with user verification or password+MFA; never OTP-first with required MFA	S21
Event fan-out	EventBridge → SNS to many channels; shared by U1 reminders and U2 alerts	S21, S23
Claim check (notification without payload)	Push says "something is ready"; the client fetches PHI over the authenticated API	S23
Tools not embeddings	Expose patient data to the assistant as typed MCP tools, never as embedded raw data	S22, S29, S43
Hot/cold store	InfluxDB for recent, S3 Tables (Iceberg) for history, one Flink job between	S27
Stream processing with keyed state	Per-patient thresholds and Random Cut Forest inside Flink	S27, S29
Edge gateway (phone as aggregator)	The phone reads on-device health stores and BLE, forwards over MQTT with its own identity	S28
Edge scoring	Quantised ONNX model as a Greengrass v2 component	S29
Human-in-the-loop approval	Step Functions holds the draft until a clinician approves	S34
Provenance / citation	Every generated sentence carries a timestamp citation to the source	S34
Option pair with a switch condition	Two valid AWS paths; choose one, write the condition for switching	S35
Evaluation harness	Measure whether a trained model beats a prompt before shipping it	S36, S43
Confidence-gated review queue	Low-confidence extractions route to a person	S41
De-identification stage	DetectPHI + Macie + redaction UI, with the Safe Harbor caveat stated	S41, S48
Link + fork composition	Link the maintained deployable, fork what you change, build only your layer	S42
Governed lakehouse	S3 Tables + Glue + Athena + Lake Formation column-level access	S48
EMPI via entity resolution	AWS Entity Resolution over HealthLake identifiers	S48
Pipelines spine / Model registry	SageMaker Pipelines + Registry + Feature Store + MLflow deployed once	S22, S50, S68
Explainability in the pipeline	SHAP + bias metrics logged to MLflow (Clarify successor)	S50
Shared engagement channel	Connect Customer campaigns serve both U1 reminders and U6	S57
Docs-as-code / content contract	Repo-owned README, ADRs, metadata; portal aggregates at build time	S61
Architecture Decision Records	MADR 4 file per lead decision; immutable once accepted	S63
Negative test as evidence	Show the wrong control failing next to the right one succeeding	S70

Appendix C — Citation map

Seed for sources.json. Each claim id links to its primary source; full wording and validator notes are in research/aws-demo-catalog/dossier.md §3.

id	claim	primary source	label
C1	LZA + LZA-for-Healthcare is organisation-only; ~US$430/month	docs.aws.amazon.com/solutions/latest/landing-zone-accelerator-on-aws/mandatory-accounts.html · github.com/aws-samples/landing-zone-accelerator-on-aws-for-healthcare	H
C2	No public single-account HIPAA baseline	github.com/aws-samples/deploy-conformance-pack-for-aws-startup-security-baseline	H (absence)
C3	Control Tower current; Organization required	aws.amazon.com/about-aws/whats-new/2026/07/aws-control-tower-account/	H
C4	Config HIPAA conformance pack	docs.aws.amazon.com/config/latest/developerguide/operational-best-practices-for-hipaa_security.html	H
C5	Security Hub CSPM standards; no HIPAA-named standard	docs.aws.amazon.com/securityhub/latest/userguide/standards-reference.html	H
C6	Macie 12 PHI identifiers, none default	docs.aws.amazon.com/macie/latest/user/mdis-reference.html	H
C7	BAA in AWS Artifact, account or org scope	docs.aws.amazon.com/artifact/latest/ug/accept-org-agreement.html	H
C8	New sign-up "not designed for regulated workloads"	docs.aws.amazon.com/bedrock/latest/userguide/compliance-validation.html	H
C9	HIPAA whitepaper archived; Healthcare Lens 2022-11-17	docs.aws.amazon.com/whitepapers/latest/architecting-hipaa-security-and-compliance-on-aws/document-revisions.html	H
C10	SRA examples CC-BY-SA-4.0 — link only	github.com/aws-samples/aws-security-reference-architecture-examples	H
C11	HIPAA Eligible Services Reference (2026-09-03); feature rule	aws.amazon.com/compliance/hipaa-eligible-services-reference/	H / ML
C12	Eligibility list is not a lifecycle signal	dossier §3	H
C13	No AWS-official U1 Tier 1	dossier §3	H
C14	U1 component samples	github.com/aws-samples/sample-intelligent-security-for-healthcare-apis · github.com/aws-samples/aws-healthlake-smart-on-fhir	H
C15	Amplify Gen 1 EOL 2027-05-01; Gen 2 versions	github.com/aws-amplify/amplify-cli/issues/14881	H
C16	Flutter lacks passkeys and AppSync Events	docs.amplify.aws/swift/start/migrate-to-gen2/feature-matrix/	H
C17	Classic Mobile SDKs EOS 2026-08-01	docs.aws.amazon.com/iot/latest/developerguide/iot-sdks.html	H
C18	API Gateway vs AppSync documented split	docs.aws.amazon.com/prescriptive-guidance/latest/modernization-integrating-microservices/appsync-api-gateway.html	M
C19	Cognito passkey / MFA rules	docs.aws.amazon.com/cognito/latest/developerguide/amazon-cognito-user-pools-authentication-flow-methods.html	H
C20	Pinpoint EOS 2026-10-30	docs.aws.amazon.com/pinpoint/latest/userguide/migrate.html	H
C21	Pinpoint successors; SNS mobile push current	docs.aws.amazon.com/sns/latest/dg/sns-mobile-application-as-subscriber.html · docs.aws.amazon.com/push-notifications/latest/userguide/reference-send-message.html	H
C22	U1 Tier 2 scaffolding; public-sector blueprint	aws.amazon.com/blogs/publicsector/architecting-hipaa-compliant-ai-agents-to-safeguard-health-data-with-aws/ · github.com/aws-samples/amazon-bedrock-agents-healthcare-lifesciences	H
C23	SageMaker Pipelines CDK spine	github.com/aws-samples/amazon-sagemaker-pipeline-deploy-manage-100x-models-python-cdk	H
C24	HealthLake regions and US$0.27/hr	aws.amazon.com/healthlake/pricing/ · aws.amazon.com/healthlake/faqs/	H
C25	Device Farm us-west-2 only	docs.aws.amazon.com/general/latest/gr/devicefarm.html	H
C26	No end-to-end RPM example; no HealthKit/Health Connect guidance	dossier §3	H
C27	U2 component samples	github.com/aws-samples/aws-appsync-iot-core-realtime-dashboard · github.com/aws-solutions-library-samples/guidance-for-aws-iot-greengrass-foundations · github.com/aws-samples/flink-keyed-random-cut-forest-example	H
C28	No IoT Core → InfluxDB IaC sample; InfluxDB not listed by name	aws.amazon.com/compliance/hipaa-eligible-services-reference/	H / ML
C29	Phone → IoT Core via Cognito identity + IoT policy; IoT Device SDK for Swift GA 2026-06-24	docs.aws.amazon.com/iot/latest/developerguide/cognito-identities.html · aws.amazon.com/about-aws/whats-new/2026/06/aws-iot-device-sdk-swift/	H
C30	HealthKit / Health Connect on-device; Apple 5.1.3	developer.apple.com/app-store/review/guidelines/ · developer.android.com/health-and-fitness/guides/health-connect	H
C31	BLE Swift sample: CDK current, iOS layer EOS	github.com/aws-samples/aws-amplify-cdk-iot-ble-swift-app	H
C32	U2 Tier 2: no example	dossier §3	H
C33	Edge Manager gone; ONNX + Greengrass v2	docs.aws.amazon.com/sagemaker/latest/dg/edge-eol.html	H
C34	Flink → SageMaker endpoint pattern (generic)	github.com/aws-samples/amazon-sagemaker-feature-store-streaming-inference-msk-kda	M
C35	HealthScribe us-east-1 only; US$0.10/min	docs.aws.amazon.com/transcribe/latest/dg/health-scribe.html · aws.amazon.com/healthscribe/pricing/	H
C36	Transcribe Medical / Comprehend Medical quiet, not retired	dossier §3	H
C37	Connect Health GA 2026-03-05	aws.amazon.com/about-aws/whats-new/2026/03/amazon-connect-health-agentic-ai-healthcare/	H
C38	Connect Health samples on Bedrock Agents classic	github.com/aws-samples/sample-amazon-connect-health-unified-clinical-workflow	H
C39	HealthScribe CDK TS sample	github.com/aws-samples/sample-healthscribe-bedrock-clinical-analysis	H
C40	Diagnosis-codes Guidance; only the Comprehend Medical slice is in code	docs.aws.amazon.com/solutions/identifying-diagnosis-codes-from-clinical-notes-on-aws/ · github.com/aws-solutions-library-samples/guidance-for-identifying-diagnosis-codes-from-clinical-notes-on-aws	H
C41	`aws-healthscribe-demo` is Gen 1	github.com/aws-samples/aws-healthscribe-demo	H
C42	U3 Tier 3: evaluation notebooks only	dossier §3	H
C43	DetectPHI does not redact; not Safe Harbor	docs.aws.amazon.com/comprehend-medical/latest/dev/textanalysis-phi.html · aws.amazon.com/comprehend/medical/faqs/	H
C44	GenAI IDP Accelerator	github.com/aws-solutions-library-samples/accelerated-intelligent-document-processing-on-aws	H
C45	IDP Guidance page; repo licence unconfirmed	docs.aws.amazon.com/solutions/intelligent-document-processing-on-aws/ · github.com/aws-samples/aws-ai-intelligent-document-processing	H / M
C46	`sample-healthcare-agents`	github.com/aws-samples/sample-healthcare-agents	H
C47	BDA default; Textract on the list by name	dossier §3	H / ML
C48	BDA sample uses A2I + Ground Truth — do not link	dossier §3	H
C49	`aws-ai-phi-deidentification`; Clario blog	github.com/aws-samples/aws-ai-phi-deidentification · aws.amazon.com/blogs/architecture/how-clario-automates-phi-pii-detection-in-dicom-images-using-amazon-bedrock/	H / M
C50	U4 Tier 3: none	dossier §3	H
C51	U4 best-covered item	dossier §3	H
C52	PDF-to-HealthLake Guidance is architecture-only	aws.amazon.com/solutions/guidance/ingesting-pdf-and-image-files-to-aws-healthlake/	M
C53	Multi-Modal Data Analysis Guidance	github.com/aws-solutions-library-samples/guidance-for-multi-modal-data-analysis-with-aws-health-and-ml-services	H
C54	Patient Entity Resolution Guidance	github.com/aws-solutions-library-samples/guidance-for-patient-entity-resolution-with-aws-healthlake	H
C55	S3 Tables not listed by name	aws.amazon.com/compliance/hipaa-eligible-services-reference/	H / ML
C56	Amazon Quick listed; artefacts say QuickSight	dossier §3	H
C57	Readmission samples 2021–2022	dossier §3	H
C58	HealthLake text-to-SQL snippet	github.com/aws-samples/extract-medical-insights-from-amazon-healthlake-with-bedrock	H
C59	Clarify closed; SHAP + MLflow successor	docs.aws.amazon.com/sagemaker/latest/dg/clarify-availability-change.html	H
C60	Connect Health capabilities and pricing	aws.amazon.com/products/connect/health/	H
C61	Amazon Connect and Lex HIPAA-listed	aws.amazon.com/compliance/hipaa-eligible-services-reference/	H
C62	Connect voice sample on Agents classic	dossier §3	H
C63	Nova Sonic AgentCore call-centre sample (MIT)	github.com/aws-samples/sample-Nova-Sonic-AgentCore-Healthcare-Call-Center	H
C64	No healthcare Connect workshop; forecasting is a product feature	dossier §3	H
C65	U1 and U6 share Connect Customer campaigns	dossier §3	M
C66	Workshop Studio `contentspec.yaml` contract	github.com/aws-samples/rancher-on-aws-workshop/blob/main/contentspec.yaml · aws.amazon.com/about-aws/whats-new/2025/11/workshops-available-aws-builder-center	H
C67	Workshop Studio pages client-rendered	dossier §3	H
C68	Guidance page anatomy; usage disclaimer	docs.aws.amazon.com/solutions/guidance-disclaimers/	H
C69	aws-samples README contract; `_pattern-model/example-pattern.json`	github.com/aws-samples/serverless-patterns/tree/main/_pattern-model	H
C70	Repo-generated site precedents; MADR 4	github.com/aws-samples/eks-workshop-v2 · aws-samples.github.io/amazon-bedrock-agents-healthcare-lifesciences/ · adr.github.io/madr/	H
C71	Comparator portals	dossier §3	M
C72	Bedrock does not train on customer content	aws.amazon.com/bedrock/faqs/	H
C73	Retention modes; Fable 5.x `aws_review`	docs.aws.amazon.com/bedrock/latest/userguide/data-retention.html	H
C74	Guardrails do not evaluate tool I/O	docs.aws.amazon.com/bedrock/latest/userguide/guardrails-sensitive-filters.html	H
C75	AgentCore Gateway Policy + interceptors	aws.amazon.com/blogs/machine-learning/secure-ai-agents-with-policy-and-lambda-interceptors-in-amazon-bedrock-agentcore-gateway/	H
C76	Security Hub CSPM AI Security Best Practices	docs.aws.amazon.com/securityhub/latest/userguide/standards-reference.html	H
C77	Retired-service dependents (dossier §7)	dossier §3	H
C78	us-east-1 primary; HealthScribe the binding constraint	dossier §3	H / M
C79	Licence census	dossier §3	H