Start with provenance
A report should name the models, observation region, task profile, time window and sample count. A chart without these details is hard to interpret. We keep reports unpublished until their numerical statements can be reconstructed from validated observations. Demonstration fixtures are never eligible evidence.
Check what is missing
Ask how failures, missing usage and partial streams were counted. Check whether a report describes one account or makes an unsupported claim about everybody. Provider status statements and probe results are different evidence types and should be labeled separately.
Look for reproducibility
The summary dataset should reconcile with every numeric sentence and chart. Derived values need an aggregation version and documented exclusions. Immutable raw records make corrections reviewable; silently replacing old observations makes it impossible to understand why a result changed.
Separate observations from advice
Speed does not establish model quality, future availability or a guaranteed best work time. Reports distinguish description from prediction and state when evidence is limited. Advertising must not dictate a ranking. This release contains explanatory guides and an empty report index, not invented benchmark stories. When measurements become available, readers should be able to identify the exact window and understand every exclusion before relying on a headline.