ONE BOUNDED GPT-6 LUNA AUTH CANARY · OBSERVED
Request Boundary: Why a Short Reply Can Follow a Large Input
A long-scroll visual companion to the article. It summarizes the article’s evidence boundary and method; it does not extend the underlying result.
At a glance
Evidence in scope
- Input: 13,289 tokens
- Cached input: 0
- Output: 10 tokens
Evidence boundary: One bounded run only. The provider total does not attribute tokens to wrappers, tools, history, or any other component.
How to use the diagram
Preserve provider-reported input, cached-input, and output fields separately; bind them to one request and version.
What this visual explains: Measure the serialized request at the provider boundary. A short visible answer does not reveal the size or composition of the request.
What it does not establish
No component-level token split, general cache claim, benchmark, or repeated-workload conclusion is established.