Building in public during closed beta

We are keeping the early cohort small on purpose. Reasoning products fail quietly: scores can look fine while the underlying prompts, sources, or feedback rubrics mis-train the habit we care about. A closed beta lets us watch real sessions, fix failure modes quickly, and avoid optimizing for vanity metrics that do not correlate with better judgment.

If you are in the beta, thank you. Your bug reports and “this feedback felt wrong” moments are as valuable as feature requests. We will post longer research-style notes here as we stabilize the core loop and expand the exercise library.