Monitor architectural decision health by checking lifecycle, ownership, retrieval, enforcement and exception signals separately. Flag records with missing owners, stale reviews, broken enforcement links, repeated violations, frequent exceptions or scope that no longer maps to the repository. Use the signals to prompt review, not to manufacture a false universal score.
Accepted does not mean permanently healthy
A decision can remain marked accepted while its service is renamed, owner leaves, enforcement rule disappears or implementation gradually diverges. Lifecycle status tells consumers how to treat a record; it does not prove that the surrounding control system still works.
Health monitoring tests the connections around the decision. It should distinguish a stale record from a broken check and a frequent exception from an actual violation.
Measure inspectable signals
Start with facts a team can act on. A dashboard is only useful when every warning links to the record, repository evidence and owner who can resolve it.
- No current owner or review date beyond the agreed interval.
- Declared paths, services or repositories no longer exist.
- Linked policy or architecture test is missing, disabled or never run.
- Applicable changes repeatedly omit the decision citation.
- Violations or exceptions are increasing in the same scope.
- Another active decision appears to conflict or supersede it.
Avoid the seduction of one score
A composite score can help prioritise a large catalogue, but weights encode organisational judgement. Show the underlying reasons and thresholds. A decision with no automated enforcement may be entirely healthy if its consequence is qualitative and reviewers consistently apply it.
Similarly, zero violations may mean the architecture is sound, the check is broken or the decision is never retrieved. Pair outcome metrics with control-coverage evidence.
Turn monitoring into review
Route each signal to an owner with a small set of valid outcomes: confirm, amend, supersede, retire, repair enforcement or open a time-bound exception. Preserve the review result in Git so agents and humans receive the same updated authority.