Latency is the round trip, not a decision

The number is elapsed time

The gateway can store how many milliseconds passed between the request and the response. That is a duration for that call. A large number can be a slow dependency. A small number can be a cached path. Neither one is a label from a safety classifier.

If the field is missing, write “latency not in the log.” Do not infer a block from a short reply.

Decisions have their own fields

A classifier decision, a finish reason, and an empty body are separate fields. Latency does not set them. A fast empty body is still an empty body, and the emptiness is what you cite, not the speed.

Do not compare two calls as “more suspicious” because one was slower. The duration is not a rank of risk.

What to file

File the response id and the latency with the unit. File any decision from its own field. Do not write “blocked” or “refused” in the latency cell.

Worked case: 80 ms read as a refusal

Response r-730 returned in 80 ms with a one-line hours answer and finish reason stop. A draft says the assistant refused the question because the call was too fast to have thought.

The note quotes 80 ms as the round trip and stop as the finish reason. The body is the hours line, which is present. Speed is not a refusal. The refusal sentence is removed.

response: r-730
latency: 80 ms
finish reason: stop
body: one hours line, present
write: duration and finish reason
not: refused because it was fast