The headline index is a rounded, provisional summary of production structural measurements. Lower is better. It is neither a percentage of bad code nor a quality label. A missing scored dimension withholds the headline; incomplete advisory evidence remains visible without being silently counted as zero.
Terminal and Markdown reports put absolute burden beside density. Each observed density shows its numerator and denominator. This matters when a clean addition changes the denominator while an existing problem remains. The score-contribution table traces integer points to raw metric IDs, and the review panel names any current density term at its formula saturation point. Top located contributors show one production finding per scored structural kind; the full bounded hotspot list follows.
The graph panel distinguishes the scored production-induced cycle graph from workspace-wide edge observations. It reports runtime versus type-only cycle groups, type-only/deferred/conditional local relationships, unresolved edges and unsupported files with their scopes labeled. These counts do not imply that runtime-selected imports were fully observed.
Documentation warnings have their own advisory section with observed block
size, effective line and word thresholds, and locations. They do not appear
among scored structural hotspots and are not a reason to delete useful API
contracts. Executable-scope and nesting findings remain separate review
leads. Operators can inspect them and set explicit metric budgets or
new-finding policy gates in trellis.yaml.
A comparable comparison carries exact, unrounded formula points per dimension alongside the apportioned integer points. It lists changed density denominators as numerator/denominator pairs and numeric changes on persistent findings. When the index is unchanged but raw measurements move, the report says so explicitly. A native source-snapshot fingerprint indicates whether selected source content changed. If that fingerprint is unavailable, the comparison says it cannot determine source change. Analyzer and scoring version mismatches remain incompatible, and configuration caveats remain visible; the source fingerprint alone cannot prove that every configuration setting matched.
The JSON comparison includes these same structured explanation fields. CLI, SDK, fleet and history use the same core report and comparison paths. Use saved baselines only when their measurement semantics are compatible; make a fresh baseline after scored version changes. Review located findings and set workspace-specific budgets rather than applying an arbitrary global cutoff.