How ranking is tested. No guesswork.
Replore's order is a fixed formula, not a model's opinion: how well a repo matches your words (its place in GitHub's best-match results), how many people use it (stars, on a log scale) and how freely you can use its licence. The same input always gives the same order, on the web page and in the plugin.
Every change must pass the goal cases below before it ships. They run in CI on every push, with
python evals/run.py. Each one names the repo a builder would actually want at the top.
Fit ranking
Relevance, popularity and licence freedom together.
Words and meaning together
When the meaning-based search is on, the fit order is blended with how well each repo's text matches your whole description and how close it is in meaning. These cases pin that blend, including the edge cases.
Couldn't load the cases. They are in evals/ on GitHub.