How Deslop ranks duplicate-code clusters
An unranked list of duplicate clusters leaves the user to decide where to begin. Deslop assigns each cluster an impact weight and sorts the report from highest to lowest.
The formula
weight = clone_node_count × (cluster_size − 1) × log2(1 + spanned_bytes)
Implemented in crates/deslop-core/src/cluster.rs::rank_weight. Three factors, all multiplicative, with one logarithmic damper.
clone_node_count — the AST node count of the duplicated fragment. Higher node counts represent larger structural fragments.
cluster_size − 1 — the number of additional members beyond the first. Two copies counts as one duplicate pair. Five copies counts as four. A singleton cluster scores zero by construction, which is the mathematically honest version of "one occurrence isn't a duplicate."
log2(1 + spanned_bytes) — the byte span, dampened by log2 so very large ranges do not dominate the ranking. Deslop addresses occurrences by [byte_start, byte_end); line numbers are derived for display.
The score increases with fragment size, additional copies, and byte span. The logarithm makes span grow more slowly than the other two factors.
What the formula deliberately excludes
- Language weight. An identical-code C# duplicate and an identical-code Rust duplicate score identically if their nodes × (size − 1) × log spans match. Language preferences belong in configuration, not the ranking.
- Signal weight. The ranking does not multiply by
embedding_cosorstructural. Those signals gate whether a cluster exists at all (the fused threshold sits at 0.85 inpair.rs). Once accepted, every cluster is ranked on the same scale. - File age / churn. File history is not part of the formula. Ranking uses the current report's structural measurements.