SecureLLMFirmin Labs
Back to summary
References & Sources

References & Sources

Companion to: Executive Summary · Technical Implementation Blueprint · Alternative Platform (OpenAI) Analysis Purpose: The authoritative reference material used to build this proposal. Each entry notes what it supports so it can be published as a "learn more" link beside the relevant claim.

How to use this list. The Primary sources section is the vendor documentation and announcements actually used to verify the architecture, configuration keys, and product behavior in these documents. The Canonical service documentation section lists the standard AWS docs a reader should consult for depth on each infrastructure component — stable reference pages, not specific claims. The Note on figures explains where the cost and pricing numbers come from (they are planning inputs, not quoted from a source).

Verification note. Links reflect vendor documentation as of August 2026. Vendors reorganize documentation and change pricing and model availability, so verify each URL and any figure in the live console before publishing to a public site.

Primary sources (used to verify this proposal)

Claude Desktop on Amazon Bedrock — deployment

  • Deploy Claude Desktop on 3P with Amazon Bedrock
    Anthropic (Claude) documentation
    claude.com/docs/cowork/3p/bedrock

    Supports: running the native Claude desktop app in third-party inference mode against AWS Bedrock; the inferenceProvider: "bedrock" setting; region pinning via inferenceBedrockRegion; the inferenceModels list format; authentication paths (bearer token, in-app AWS sign-in via IAM Identity Center, named profile); per-region Bedrock model access.

  • Deploy with MDM
    Anthropic (Claude) documentation
    claude.com/docs/third-party/claude-desktop/mdm

    Supports: the Windows machine-policy registry path HKLM\SOFTWARE\Policies\Claude; the REG_SZ value constraint and the "values directly under the key, not in a subkey" rule; the macOS Managed Preferences path /Library/Managed Preferences/com.anthropic.claudefordesktop.plist; .mobileconfig / .reg export; the "deploy configuration first, then the app" sequence; managed configuration making the in-app window read-only.

  • Configuration reference
    Anthropic (Claude) documentation
    claude.com/docs/third-party/claude-desktop/configuration

    Supports: the feature-lockdown keys (userPluginMarketplacesEnabled, userPluginUploadsEnabled, skillCreationEnabled, disableBundledSkills, disabledBuiltinTools, disableDeploymentModeChooser); the inferenceModels array structure and labelOverride.

  • From developer desks to the whole organization: Running Claude on Amazon Bedrock
    AWS Machine Learning Blog
    aws.amazon.com/blogs/machine-learning/…running-claude-cowork-in-amazon-bedrock

    Supports: the enterprise deployment pattern; VPC endpoint / network isolation; in-region and cross-region inference profiles; OpenTelemetry export to CloudWatch and CloudTrail audit logging; consolidated AWS billing and cost attribution.

  • Claude in Amazon Bedrock
    Claude Platform documentation
    platform.claude.com/docs/en/build-with-claude/claude-in-amazon-bedrock

    Supports: Claude model availability and invocation on AWS Bedrock; model identifiers.

OpenAI models on Amazon Bedrock (alternative platform)

Data residency & cross-region inference (the residency caveat)

Pricing basis

  • Amazon Bedrock Pricing
    AWS
    aws.amazon.com/bedrock/pricing/

    Supports: the on-demand, per-token billing model that underlies the TCO. The specific per-1M-token rates and the ca-central-1 figures in these documents must be confirmed here in the live console — see the Note on figures below.

Canonical service documentation (for depth on each component)

Stable AWS reference pages for the infrastructure components. Link these where a reader wants to understand a specific control in depth.

Note on figures (important for accuracy on a public page)

The cost, pricing, and forecast numbers in these documents are planning inputs supplied for modeling — not quoted from a vendor source:

  • The per-1M-token rates ($3.00 input / $15.00 output for Claude; $2.50 / $10.00 for OpenAI GPT-5.6), the ≈$110/month operating overhead, and the ≈$43,320 / ≈$35,520 annual forecasts are estimates used to size the deployment. On-demand model inference is consumption-billed and therefore variable.
  • Do not attribute these figures to AWS, Anthropic, or OpenAI on the public page. Present them as the organization's own planning estimates, and confirm live rates and ca-central-1 model availability against the Amazon Bedrock Pricing page and the AWS console before publishing.
  • The specific model identifier ca.anthropic.claude-3-5-sonnet-20241022-v2:0 and its Canada-contained residency behavior must be verified in the console; if it is not strictly contained to ca-central-1, the direct in-region model identifier should be used instead.
Back to summary