Rendered directly from the canonical source.
This Governance audit examines architecture drift, recurring bypass, exception growth, and the long-lived health of boundaries and guards.
Route
Declare the audit scope, selected indicators, evidence window, exception baseline, and negative-control evidence. Evidence accounting is complete when every in-scope indicator has grounded evidence, an explicitly unavailable source, or a reasoned exclusion.
Apply SKILL.md’s Closed/Open contract to the audit. Unavailable sources or exclusions that could change a load-bearing decision contribute applicable findings or decision-changing unknowns and therefore select Open. In either state, additionally return justified no-action or subtraction candidates, indicator owners or explicit ownership gaps, the next review trigger, and any next snapshot.
Indicators
Lagging symptoms such as incidents and rewrites arrive late, so use leading indicators only where they earn their cost. Interpret level and direction against repository context: expected product growth can raise a healthy count. Reuse a repository-defined review trigger when present; otherwise a concerning trend creates a finding for its owner, not an automatic work item.
| Indicator | How to read it |
|---|---|
| Suppression count — lint disables, unchecked casts, skipped tests, baseline entries | Default the trend and baseline to shrink-only. A justified increase is a visible decision with authority, reason, narrow scope, owner, review trigger, and removal condition; unreviewed growth is exception accretion. |
| Negative-control freshness | Track the last time a planted or known violation was detected. Staleness questions the guard mechanism even when the repository is clean. |
| Observed violation history | Track real violations separately as pressure on the boundary. A long quiet period does not by itself justify retiring a falsifiable guard. |
| Public surface growth | Compare exports, fields, and options added per period with delivered behavior or outcomes. Surface outgrowing delivery means promises may be minted as a side effect. |
| Concept count | Compare new nouns—services, managers, layers, configuration keys—with retired concepts or explicitly accepted net growth. |
| Duplicate truth or stale authority | The same fact is independently editable through multiple routes, or its declared source or owner no longer matches the active change path. Deliberate federation and replicas are not defects when the applicable section 3 dimensions and reconciliation are explicit. |
| Boundary or accountability divergence | Code widens or crosses a declared boundary without corresponding contract and decision-right evidence. Internal churn behind an unchanged contract is not drift. |
| Bypass frequency | Track how often the governed path is skipped. After controlling policy is accounted for, a rising rate may expose avoidable route cost; it is not evidence of individual misconduct. |
| Undeclared cross-boundary imports | Count imports that violate declared dependency direction or bypass a public surface, not directory kinship by itself. |
| Hand-edited derived files | Diffs touching declared generated files suggest the generator costs more than a direct edit; improve that path before normalizing hand edits. |
Evidence Retention
Prefer existing version-control history, automated-check artifacts, and metrics over a new Keel-owned store. If a new snapshot is justified, declare its scope, cadence, retention, owner, and deletion condition. Do not append unbounded history or build a second source of truth merely to measure drift.
Indicators are projections of health, not health itself. Declare each scope, evidence window, and known blind spots. Validate a quiet signal before treating it as proof of health or as grounds for retirement.