Setting the file. One moment.
Subchapter 23.19
references/output-templates/recommendation-doc.mdMarkdown5 KBView on GitHub
<2–3 plain-language sentences: recommended runtime + why, for a non-technical reader.>
<bulleted summary of the answers that drove the decision; each answer carries provenance (wizard defaults vs user-stated).>
<rationale + the top scoring signals; business framing then technical specifics. Quote each score the way scoring-result.json states it — a bare number. The engine publishes no maximum, so a fraction or a percentage (“40/56”, “71%”) asserts something it never computed.
For an AgentCore verdict, name the compute type from design.json.agentcore_compute_type and what drove it: microvms — “serverless microVMs (8h sessions, consumption billing, $0 during I/O wait)”; instances — “Instances compute type: AWS-managed EC2 via a capacity provider (sessions to 14 days, GPU/instance-type choice, EC2 + management-fee billing)” and name which answer routed there (>8h duration, GPU/heavy compute, or the instance-type requirement). Carry the scoring warning about the 14-day ceiling and pricing model verbatim into this section.>
<table: unit id | workload class | verdict | trigger | one-line rationale>
<the platform decision: consolidated onto X / split, and why; the interconnect
(queue/gateway/none) in one sentence.>
<Worker deployment plan per task queue: Tier 1 choice + which rule fired + rationale (cite decision-refs/temporal.md). Execution-tier plan per Activity class: Tier 2 targets (Light-IO in-process with hygiene reminders, agent-session scored, short-tool/heavy mapped). Commercials: selected by CURRENT server state — self-hosted → Way 1 gets Marketplace subscribe flow; already-on-Cloud gets billing-unchanged note (never re-pitch); Way 2 gets self-host stack. Cutover runbook: the selected runbook WITH its preconditions (runbook 1’s graceful-drain/Activity-retry preconditions are not optional). Bedrock follow-up: if non-Bedrock LLM calls detected AND user said Yes → point to /aws-startup-advisor:llm-to-bedrock + replay safety per decision-refs/temporal.md runbook 3 + shim caution if deeply-coupled; if user said No → “Later, optional: Bedrock” section with the same replay-safety pointer.>
<INSERT the Mermaid block + ASCII fallback produced by the Generate diagram step (Plan 3) — the blocks themselves, pasted in. Not “see diagram.md”: this doc is read on its own.>
<eliminated/lower-scored runtimes and why.>
<relevant rows from the runtime service cards.>
<Identity, Observability, Guardrails, Scaling, Tool/Gateway, Protocols — from the service card.>
<final service list from confirm.json, with why each; note which are free.>
<per model-bearing unit: model display name, path-specific model ID, API path, resolved invocation/CRIS ID or TODO, joint-decision reasoning, compatibility groups, architecture impacts, [BLOCKS]/[TUNE], evaluation mode, rollout gate, catalog date, and live verification status. State that runnable account access is established only when the live probe passed. For migrate, include “see the llm-to-bedrock skill for pricing”.>
<Build paths (estimate.json exists): the band + assumptions + “order-of-magnitude” disclaimer; note that per-unit breakdown and cost drivers are detailed in estimate.json. Migrate (estimate.json exists): present the per-unit target-state bands + breakdown from estimate.json. ALSO note that precise TCO comparison and current-spend delta are produced by the migration plugins. Do NOT invent a dollar band for anything estimate.json lacks (e.g., if Estimate failed, or for add_capabilities which bypasses Estimate).>
<scaffolding pointers; handoff pointers if applicable. For AgentCore Build paths, point to the
deployment plan at plan.md (the staged idea → plan → deploy steps) and the generated POC at
poc/ — agent code + deploy.sh + README — and note the user runs ./deploy.sh themselves (it
creates real, billable AWS resources).>
<from freshness.md template: date; AWS MCP Server-verified fields; public-web-verified Temporal / AWS Marketplace facts with source URLs and verification dates when Temporal units exist; cached fields with original snapshot dates; verify disclaimer.>
<target maturity; readiness status; control gaps; tier-appropriate release and evaluation gates; recommendation status; deferred verification requirements with owners. State that a provisional recommendation or deployable POC is not private-beta/production launch approval. Include an I/O-wait billing benefit only when current-run verification records support it.>