Don't merge two failure causes under one label in status reports (e.g. "formula / port-copy class")
In a multi-agent status relay, a coverage gap was reported under a combined label, "formula / port-copy class". The human had seen port copies fixed days earlier, so they read the label as a regression. In fact there were two separate groups: - signals the file stores only as a formula, which need an evaluator; - port copies left unlinked by a deliberately conservative safety rule.
The groups have different owners, different fixes and different sizes. A review by outside experts then showed that much of the second group was policy, not missing data: one coarse rule refused a whole file's links.
Lessons: - Report each refusal reason as its own row, with counts. Never use "X / Y class". - When a human asks "didn't we already fix X?", answer with what X covers now, what remains, and why, in plain words. - Ask an outside reviewer to classify your "refused" rows against a reference reader. It quickly separates "data absent" from "our policy refuses it".