Your team needs to build an idempotent document processing pipeline using the Claude API. Each document should be processed exactly once even if the triggering event fires multiple times (e.g., due to message queue redelivery). What design ensures idempotency at the API integration layer?

The correct design is to implement idempotency in the application layer: use the document ID as a deduplication key in a state store, check before calling the API, and mark documents as processed atomically after a successful response. The Anthropic API does not natively support idempotency keys for deduplication, so the check-before-call plus atomic mark-as-processed pattern is the standard approach for building idempotent event-driven pipelines. The design that uses an idempotency_key parameter in every API request with a deterministic key derived from the document ID and claims the API will deduplicate requests with the same key within a 24-hour window is wrong because Anthropic's API does not expose an idempotency_key request parameter; this feature does not exist in the API as described. The design that relies on Claude's inherent determinism by setting temperature to 0 is wrong because while temperature 0 produces more consistent outputs, Claude is not guaranteed to produce bit-identical responses on identical inputs, and harmlessness of duplicate processing depends entirely on the downstream system — duplicate writes to a database are not harmless. The design that uses Claude's streaming API and its unique stream ID per request to detect and discard duplicate responses is wrong because streaming stream IDs are connection artifacts for a single request, not deduplication mechanisms across requests.