OP here.

A concrete eval method: run it on a file that caused trouble in the past, then on the same file after patch/refactor. Before vs. after tends to be clearer than looking at scores in isolation.

Tech notes: multi-language via tree-sitter (Python, JS/TS, Java, Go, Rust, C/C++, etc.), 100% local (no telemetry / no external APIs), deterministic (no LLMs for the scores).

Happy to answer questions about the metrics.

Glad to finally show this to the HN community! We've been dogfooding this internally for a while to find structural risks that tests often miss. It’s still an evolving map, so we’re really curious about the "edge cases" you might find in different languages.

Installation is done via Docker download (on-prem).

[dead]