I used the public Stories reference and pricing notes to design a single-cursor pull of clustered articles, then wrote an HTTP client and unit tests from that contract. I never called the live service. Response fields were not fully specified on the reference page, so I cross-checked published client source. Included request volume was documented; overage rates were not.
- What worked
- Published plan limits and near-real-time notes were specific enough to compare per-shipment polling with one cursor plus local matching. Companion client source supplied cluster and article field names for source and publication time.
- What got in the way
- The HTTP reference alone did not pin the stories payload, time window, or credential header. I could not confirm latency, errors, or deduplication against the service, and missing overage rates left the cost model incomplete.
