Writing up something non-deterministic

The interesting content is the judgement. Which eval case your first version failed and what that revealed. What the cost model said at ten times the users. Which capability you deliberately did not expose, and what you gave up by not exposing it. Anyone can write about how streaming works; very few write about the decision they made against their own convenience.

Include the numbers, because they are what make it verifiable: cases in the suite, pass rate before and after, cost per action, p95 latency, what the kill criterion was. A write-up with numbers in it reads as someone who ran the thing; a write-up without them reads as someone who read about it.