Recent Work

Problem

A language only starts to feel like yours when you speak it, but a conversation partner at your level, with time and patience, is hard to find - and the exercises in any textbook or app run out long before the learning does. The same gap shows up before an exam like the DTZ: official practice papers are used up fast, and learners end up memorising answers instead of training listening and reading.

Architecture

• The verifier solves, never validates. Shown the answer key and asked whether it holds, the model rationalized broken items into passing. Asked instead to answer the question blind and list every supported option, it caught each one; the verdict then follows mechanically from what it picked.
• Conservative in one direction only. A false rejection destroys a sound item while a false acceptance merely preserves the status quo, so every error path - unreachable provider, malformed reply, out-of-range answer - returns a pass, and rejection requires both a named defect and high stated confidence.
• Constraints belong in code. A prompt line capping how often one grammatical category may repeat across gaps was advisory, and was quietly ignored. The model now declares a category per gap and the server rejects the whole sample when the spread is too narrow.
• Timeouts that already landed. A read timeout means the request reached the engine and is still generating; retrying it starts a duplicate job on a box that handles one at a time. Those failures now fail fast and alone, with generation serialized instead.
• Forked screens over shared branches. Adding a second exam format by branching the working one put a format already in learners' hands at risk, so the new one got its own service and screen tree behind an explicit never-touch list, verified by diff.