A second small follow-up to 0.17.0, from the same day's measurement runs of the Tutor and the Rich check-up.
One-word answers get a hint in your language. Articles, pronouns and short particles cannot be explained in the language you are learning without using the word itself, so the Tutor used to fall silent on exactly those items. It now explains that step in your native language, says so in its first sentence, and still never names the answer.
Rich check-ups generate more reliably. Where the model put the correct option in the wrong place, the set is now corrected instead of being thrown away and generated again, so fewer sittings ask you to retry.