Evaluation

eval/ holds ten Astro tasks of increasing size (a component, a utility, pages, a content collection, API routes, an island, i18n routes, a blog) with hidden acceptance tests the harness never sees while it works.

ruby eval/run_eval.rb --config config.json --profile-map eval/profile-map.example.json

It prints one table per profile mapping: solved, solved without fix rounds, regressions, findings confirmed, fix rounds, tokens and time, plus tokens and time per role. Preparing the starting repository is described in eval/README.md.