Field note
Draft readings without softening the findings
How we run draft reviews so technical contacts can correct facts without sanding consequence out of a reliability report.
A draft reliability report that cannot be challenged is propaganda. A draft that becomes a negotiation over every uncomfortable sentence is also useless.
We invite the technical contact to mark factual errors: wrong node counts, misread deploy windows, a dependency we attributed to the wrong team. We do not invite them to replace “retry storm” with “occasional elevated error rates” when the analytics show otherwise.
A rule that keeps readings short
Disputed interpretations go to the owner reading with both views labelled. The final report still chooses a primary reading and states the residual uncertainty. Cluster reliability reporting is not consensus theatre; it is a responsible claim about risk, backed by the evidence the engagement was allowed to see.
When access was incomplete
If runbooks were withheld or metrics exports truncated, the report says so in the limitations section. Pretending completeness is a faster way to lose trust than admitting a blind corner.