Zaka

Zaka certifies JLPT question difficulty empirically instead of asserting it. A governed pipeline retrieves grounded vocabulary, generates a question, gates it deterministically, audits it, then tests it live on four simulated JLPT learner levels (N5-N2) — if the difficulty gradient is wrong, it retrieves again and retries under policy before certifying. Built for the Vultr track: every model call runs on Vultr Serverless Inference.