Bedrock routing without the router Lambda: one state machine, two model calls per question

The toll for classification is smaller than it looks, and that is the whole case for the design. The classifier state is capped at 10 output tokens, while the Claude Haiku branch is allowed 500 ^[4][12], a ratio of fifty to one on the output side ^[16]. Breakeven is therefore not close: if the classifier correctly diverts even a modest slice of traffic off Haiku and onto Nova Micro, the fixed cost of asking "simple, code, or creative" is recovered ^[4]. The arithmetic only turns bad if the classification is wrong often enough that answers have to be retried by a human, and the walkthrough offers no measurement of how often that happens.

The second thing worth pricing is what "zero application code" actually removes ^[1]. The routing logic is a JSON document, but it is not free of programming: `States.Format` assembles the classification prompt inside the state machine from the user's question ^[9], and a `ResultSelector` plus `ResultPath` pair narrows the classifier's response to `$.classification.category` while keeping the original `$.question` alive for the answering states ^[10]. The author flags the trap directly: omit `ResultPath` and the classifier's one-word answer overwrites the question you were trying to route ^[10]. That is a bug class you get in a state machine and rarely in a router function, where the request object usually survives by default. The logic did not go away, it changed notation, and the notation has no unit test harness attached.

What did genuinely leave the request path is the compute. The optimized `arn:aws:states:::bedrock:invokeModel` integration lets each Task call the model directly ^[6], so there is no function to package, no runtime to upgrade, and no cold start between the classifier and the answer. The state machine defines four Bedrock invocations across four states, one classifier and three answerers ^[15], and a single execution touches two of them ^[13].

The maintenance did not leave either, it moved. Model identifiers on Bedrock change often enough that the post tells you to check `aws bedrock list-inference-profiles` before running anything, and notes its own IDs were only current at writing ^[8]. The dated string in the definition is the Haiku profile ^[5], which means the branch most likely to rot is the expensive one, and the failure lands on exactly the requests you routed there because the small model was not good enough for them. A JSON document in a console does not warn you about that; a dependency file in a repository sometimes does.