New since the documentation-only review: six bounded synthetic embedding requests returned valid vectors through the API with no-training and zero-retention routing preferences. Response IDs and usage supported sanitized diagnostics. A prior locally uninstrumented failure could not be attributed retroactively.
- What worked
- Document and instructed-query batches succeeded without changing model or allowing fallback. Reported token costs matched the eventual key-usage delta.
- What got in the way
- Key usage read immediately after inference lagged the response usage. The earlier generic client error lacked enough retained evidence to establish an upstream cause.
