Chains

Chain composes three chainlets: transcribe (audio→text), summarize (text→summary), translate (summary→localized). Operator returns dict from transcribe instead of the declared TranscribeOutput pydantic model.

Each chainlet declares typed inputs and outputs (pydantic models). The Chains runtime validates at the hop boundary; returning an untyped dict triggers a schema-mismatch failure at the next hop. Declare the model and import it from a shared package consumed by both chainlets.

Finance asks: how much of the chain's spend is the LLM hop vs the embedding hop?

Chains usage reporting breaks GPU-seconds out per chainlet. Aggregate over the billing window per chainlet name to attribute cost. Do not divide chain-level spend by chainlet count — costs are not uniform across hops.

Operator changes the TranscribeOutput schema (adds a `language` field). Summarize chainlet still references the old shape.

Treat the input/output schemas as a shared contract. Bump both chainlets in lockstep, push them as a single chain deployment, and avoid in-place schema edits to a production chain. Use additive changes (optional field with default) when full lockstep is not feasible.