A teaching system should distinguish that a learner liked a reply from what they could retrieve later. Keating 2.2 starts making that distinction durable. It keeps the original learning events, then derives cautious beliefs from them instead of collapsing a session into a thumbs-up or thumbs-down.
Evidence-led learning
Quiz results, structured diagnostic answers, and flashcard reviews now converge in a topic-level learner profile. The profile tracks mastery, retention, confidence, strengths, weaknesses, and the learner's reported challenges. Performance carries more weight than sentiment; a single excellent quiz remains provisional, while repeated retrieval over time can become real evidence of retention.
Matching diagnostics score immediately when they have an answer key. Free-text answers are saved as pending checks and can be judged by the tutor as correct, partial, or incorrect, with a misconception note only when the response supports one. That keeps the record useful without pretending every learning question has a simple right answer.
Feedback with a referent
Message feedback is now an attribution record. It saves the stable assistant-message id, cleaned generated response, timestamp, session id, comment, and a clearly labeled session-title topic context. Later analysis can therefore inspect the answer the learner actually reacted to instead of guessing from whichever topic happened to be current.
The session title itself is also calmer: automatic AI titling waits for two completed learner–assistant exchanges, giving it enough material to name the work rather than echoing the opening request.
Browser runtime and identity
This release also strengthens the browser-side operating surface: safer local-execution and project-file routes for the agent runtime, richer per-model benchmark and reward attribution, hardened Hyperframes iframe controls, and repaired OAuth callback and token paths for OpenAI Codex.
Verification
Keating 2.2 ships with focused learner-profile, diagnostic, OAuth, agent-runtime, project-file, benchmark, and Hyperframes coverage, plus the normal root and web type/build checks used for release preparation.