Where your data goes

This is Amnetic's public model-layer disclosure for marketplace inference. It is derived from the canonical ZDR inference posture and was verified for this PR on 2026-09-08. AWS Bedrock is the inference substrate; Anthropic is the developer of the Claude models AWS serves and does not receive content from this pre-purchase marketplace inference.

Before purchase, no model provider receives your content. Marketplace inference runs on AWS Bedrock in Amnetic's AWS account; under AWS Service Terms 50.12.5 AWS may share model-usage information with a model's provider but not Your Content, and Amnetic denies the provider_data_share retention mode and never opts into Anthropic abuse-detection content transfer.

This sentence is scoped to model-provider handling before purchase. After purchase, buyers may send purchased content anywhere by design; Amnetic cannot control how a buyer handles content it has bought.

Models and retention

The marketplace routes the following Claude models through serverless AWS Bedrock in Amnetic's AWS account. The region column is the current US geography destination set, not a promise that every request stays in one region.

Model Substrate US geography destination set Retention Training
claude-opus-4-6 AWS Bedrock us-west-2, us-east-1, us-east-2 AWS-only abuse-detection retention of up to 30 days for models AWS lists (none we route as of 2026-09-08); prompt-cache prefixes up to 1 hour AWS-hosted Anthropic-on-Bedrock terms bar training on Customer Content
claude-sonnet-4-6 AWS Bedrock us-west-2, us-east-1, us-east-2 AWS-only abuse-detection retention of up to 30 days for models AWS lists (none we route as of 2026-09-08); prompt-cache prefixes up to 1 hour AWS-hosted Anthropic-on-Bedrock terms bar training on Customer Content
claude-haiku-4-5 AWS Bedrock us-west-2, us-east-1, us-east-2 AWS-only abuse-detection retention of up to 30 days for models AWS lists (none we route as of 2026-09-08); prompt-cache prefixes up to 1 hour AWS-hosted Anthropic-on-Bedrock terms bar training on Customer Content

The abuse-detection residual is AWS-only unless Amnetic explicitly opts into a provider-data-sharing mode; the ZDR posture denies that mode. Prompt-cache prefixes are account-isolated, stay on AWS, and are not used for training. A cached prefix can include the caller-session transcript (ADR-092). Invocation logging is disabled for marketplace inference.

The E2B sandbox boundary

E2B is Amnetic's managed sandbox provider, a separate boundary from the model provider. While a buyer agent runs, seller listing content and the caller session material needed for that evaluation transit the E2B microVM. The VM can reach only the market-read, LLM, and suggestion proxies; it has no arbitrary network egress and cannot persist content after the session ends. E2B does not receive content as a model provider, and this page's Bedrock claim does not turn into a claim that no third party ever handles seller data.

Amnetic-side retention and operator access

Sessions carrying the signed CapDebugLLMContent capability can emit full LLM request and response content to the llm-proxy CloudWatch log group through the operator debug keyhole. Only the operator-controlled session-token minting path can grant that capability; ordinary buyer agents and sellers cannot self-assert it. Access to those CloudWatch logs is limited to operator IAM roles and the internal amnetic-internal logs <session-id> path guarded by those credentials. The posture sets and asserts a CloudWatch retention of no more than 30 days for that debug content. Normal sessions do not durably archive the transient audit debug fields.

The buyer-agent playground is different: its sealed traces persist full transcripts by product design and follow the playground retention and crypto-shred design, not the Bedrock provider-retention claim. Do not upload content to that playground unless its separate retention terms are acceptable.

Related documentation