H.7.3 Evidence Selection

Selection is a risk policy, not a promise that every change has one exact test set.

• Define protected scopes and global invalidators whose evidence ordinary narrowing cannot omit.
• Represent selector errors, incompatible baselines, missing graph data, and unknown changes explicitly.
• Choose broad fallback, blocked admission, or an accountable bounded decision from the affected scope's risk.
• Validate selection against broader lanes and incidents; measure misses, over-selection, latency saved, compute, and wait.
• Use Bazel affected-target mechanics where useful without making the policy Bazel-only.