A WORD AI KEEPS FAILING NOW SAYS SO If buying a sentence failed, the button came back looking exactly as promising as the first time — and each press is two calls to the model. One word was asked five times before it worked. After two failures in a row the card now says what went wrong: that the reading over the word kept coming out wrong, or that the sentence came back without the word in it. The word is not closed for good — open the app again and the button is back. "SENTENCES" THAT WERE ONLY THE WORD ITSELF ARE GONE For a word carrying a particle — 文章を書く — the model sometimes answered with the word and a full stop, which is technically a sentence and teaches nothing: the drill built from it blanks out the whole line. Those are refused now, and any that were already saved are removed the first time you open this version. The word simply goes back to offering the button. WITH A SUBSCRIPTION, THE COUNT SAYS "THIS PERIOD" RATHER THAN "THIS MONTH" Your monthly allowance runs from the day you subscribed, not from the first of the month — so if you subscribed on the 20th, it renews on the 20th. The date shown was always right; the sentence beside it said "this month" anyway. The free allowance is unchanged: that one really is calendar months.

WORDS AI COULD NOT WRITE A SENTENCE FOR ARE OUT OF THE DEAD END If buying a sentence ever failed, the word was left with three progress bars instead of four and no way back — the message said "try again" while the button was already gone. Restarting did nothing, and restoring a backup did nothing either, because the state traveled with it. Those words are unlocked the first time you open this version, with no prompt and nothing spent. TRY AGAIN COMES BACK TO THE WORD CARD, AND THE MESSAGE SAYS WHAT ACTUALLY HAPPENED "AI could not write a sentence for this word" was sometimes untrue: the model did answer, it just wrote a sentence without the word in it, and we were the ones who turned it down. That is what it says now, and the button stays, because a second try on the same word can work. One thing is unchanged: when the model answers that this word has no ordinary sentence, you get its reason rather than another attempt. QUESTIONS ABOUT A READING NOW SHOW WHAT THE SENTENCE MEANS A sentence you cannot understand is not context — and this app promises to teach words where they stand. Both "how is this read" and typing the reading now carry the sentence translation underneath, on the same Translation setting that governs it after an answer. Questions about meaning do not get it and will not: there the translation is the answer. THE APP STARTS FIVE TIMES FASTER The 3051-word catalog was prepared on every launch, before the first frame appeared. Measured: 86% of that time went into 70 calls of one helper loop that picks wrong answers. The questions and their choices are identical to before — checked by comparing the whole catalog, entry by entry. QUESTIONS ABOUT MEANING NAME THE WORD INSTEAD OF POINTING AT IT The prompt used to read "What does the marked word mean here?", and the marking is a color. If you use VoiceOver, there was no way to tell which word was meant — the screen reader read out the whole sentence with nothing to single it out, so the question could not be answered. The prompt now says it outright: "What does 「本」 mean here?". 23 WORDS IN N5 GET A SECOND SENTENCE Nearly every word in the catalog has exactly one sentence, so a second review of the same skill showed the very same screen, character for character. The first twenty-three N5 words now have two contexts each and the rotation has something to choose from. The rest of the level follows in later releases.

QUESTIONS ABOUT A SPELLING CAN NO LONGER BE SOLVED WITHOUT READING A SINGLE CHARACTER Asked how しょうじき is written, you were shown 正直 next to 直す, 直る and 正しい — three spellings with kana in them and one without. The answer was the odd one out, and counting kana was enough to find it. Measured across the whole catalog: 854 of these questions handed out at least one variant you could reject knowing nothing, and 35 could be solved outright. Every variant now has to fit the reading on screen. NO WRONG ANSWER SAYS THE SAME THING AS THE RIGHT ONE Asked what 開く means in its sentence, you could meet both "to open (a book, one's eyes)" and "to open" on the buttons below. Two defensible answers, one of them scored as a miss. Variants that merely restate the correct meaning, more narrowly or more broadly, are now kept off the screen. MY WORDS GETS THE SEARCH FIELD AND THE PERCENTAGES THE CATALOG HAS Two lists doing the same thing looked like two different screens: the catalog searched by kanji, kana, romaji and meaning and said how much of each level was done, while your own caught words offered a count and no way to find anything. Both screens now work the same way. The percentage here counts your own words at that level, not the catalog's. THE WORD BEING ASKED ABOUT IS MARKED IN YOUR ACCENT COLOR It used to be the accent color and an underline. The underline now shows up only where color alone cannot carry the difference — that is, when Differentiate Without Color is switched on in Settings. Everyone else gets the cleaner sentence, and this is the same marking the sister apps use.

YOUR OWN WORDS AND SENTENCES CAN BE DELETED Your own entries were the one thing here you could add and never take back — a typo, a sentence written to try something out, a word that stopped being worth the time. The only way out was erasing everything in Settings, a hammer for a drawing pin. A word or a sentence now goes on its own, and before a word goes the app says in numbers what disappears with it and what stays. Answers you already gave remain: they describe what actually happened, and deleting them would rewrite history and could break a streak retroactively. THE REVIEW BUTTON STOPS BEING GRAY THE MOMENT REVIEWS COME DUE Leave the app sitting open, come back a few hours later, and the button was still gray although the words were already waiting. The app now notices the hour a review ripens instead of waiting to be launched again. WHEN AI IS UNAVAILABLE, IT SAYS WHICH OF THE THREE REASONS IT IS Under a grayed-out AI button there is now a line naming the actual reason: no connection, the service switched off on our side, or your remaining allowance could not be read. One sentence covering all three would have been wrong about two of them. THE STUDY DAY STARTS AT THE HOUR YOU PICKED — ON EVERY DAY OF THE YEAR On the night the clocks change, the boundary moved from four in the morning to five. One day a year, quietly, so nobody reported it and nobody could have. It came out of a test that ran two ways of computing the boundary against a whole year, hour by hour.

READING QUESTIONS STOP GIVING THEMSELVES AWAY A word written part in kanji and part in kana used to hand you the answer: the wrong choices differed in the kana you could already see, so the ending was enough to rule them out without knowing the character at all. That shortcut is gone. Where a question cannot be asked honestly, it is no longer asked. KANJI DRAWN THE JAPANESE WAY On a phone set to Polish or English, the system was drawing Japanese text in Chinese letterforms — close enough to miss, different enough to learn a stroke wrong. Every Japanese line in the app now asks for the Japanese shapes, whatever language your phone is in. YOU CAN SEE HOW FAR INTO A SESSION YOU ARE The progress bar left the title bar for the full width of the screen, so a single answer moves it visibly instead of by a hair. VoiceOver reads it as the question number now, not as “progress indicator”. ANSWER CARDS YOU CAN SEE IN THE LIGHT THEME In the light theme the answer cards stood on the same white as the screen behind them, and only a hairline told them apart. The screen behind them is gray now, the way the rest of the app already was.

N2 AND N1 — THE WHOLE JLPT RANGE The last two levels are here: 750 words in N2 and 900 in N1, each with its own sentences written for it. Anyone who bought Premium before this version gets both at no extra cost — the promise the store description made is kept. A LEVEL CAN BE BOUGHT ON ITS OWN Until now there was one purchase and it opened everything, including what nobody needed yet. Levels are now sold one at a time, in pairs and threes, or all at once — and buying more at a time is always cheaper than piece by piece. WORDS FROM WHAT YOU ACTUALLY READ A sentence from a manga or a textbook goes into the app: typed in, or scanned from a photo. The text in the photo is read by the phone itself, so the photo never leaves it. The words worth keeping come out of that sentence with their readings and join your reviews next to the catalog ones. A WORD CAUGHT INFLECTED GETS ITS FOURTH DRILL 読んでいた caught in a sentence cannot be asked about as a written form — there is nothing to ask about yet. One sentence written for its dictionary form adds that drill, and the word stops being a three-quarter word. AI ANALYSIS — SEPARATE, AND ONLY IF YOU ASK Pulling the words out of a caught sentence, and writing the dictionary-form sentence, is what the model does — and it is the only place where the app talks to the network. The layer is extra and entirely optional: without it the catalog, all four question types, reviews and progress count exactly the same. Every call happens on your explicit request, with the remaining count next to the button. A few calls a month are free; beyond that it is a separate subscription, which does not cover JLPT levels and never will. TWO HUNDRED GLOSSES READ AGAIN Every level went through a meaning review this release, including the ones sold long ago. Two hundred and five glosses were corrected, and fifty pairs of words that shared one gloss word for word — two cards saying the same thing — are down to none in both languages. A handful of words moved level, because a level says what the app asks about, not what the word is.

Pronunciation — a speaker button on the new-word card, on the feedback screen and on every word page. It reads the kana the catalog holds, not a guess from the kanji, and it never appears where it would give a question away. A review day has an end to it. Reviews arrived without a ceiling, so a day that fell due all at once could ask for hundreds of questions — and a single mistake brought the queue back every ten minutes until midnight. Each word now gets at most two questions while you are learning it and one once it is settled, with a ceiling of a hundred and fifty for the day. The rest waits for tomorrow instead of piling onto today. Reminders for overdue reviews. Off until you turn them on, and then only on days when something is actually due. The time is set to the minute. Nothing leaves your device. Holiday mode. Going away no longer costs you a wall of reviews: the schedule freezes, every review moves forward by the length of the break, your streak survives and reminders stay quiet. Practice still works if you feel like it. Statistics — accuracy over time, answers a day, words mastered and a reading/meaning/writing breakdown. Under Progress on the Today screen. A "How it works" guide — what decides when a word comes back, why a skip costs nothing, what the three skills on a word page mean. In Settings, under Support. Say what you think about a word: suspend it so it stops being asked, mark one you already know, keep favorites and write your own note. Progress for a single word or a whole level can be reset on its own. Targeting your weak spots is for everyone. Weighting questions towards your weakest material used to sit behind the purchase. The purchase now sells one thing only: how much material you get. The free version teaches less, not worse. Reviews and new material close the day, and nothing else does. The number on "New words" is what is left of today's intake, not everything in the catalog; material beyond that intake, or above your level, comes in with your consent and with the number, not quietly. "Mastered" now means "it survived the gap": three successful reviews spread over 21 days have been added to the threshold. The percentage on a word also weighs the three skills apart — meaning in context counts most, picking the written form least. Speed does not judge you until you ask it to. Answer time was measured always and without a word on screen, and it decided your review grade, so reading a sentence carefully cost you a shorter interval. It is now a setting, "Speed affects your score", off by default. The session is yours to set: the daily number sets your pace, the session one sets how much comes at a time. The sentence translation has three states — always, on tap, never — and it stays out of sight while you are working an answer out. The answer field opens in Japanese. Typing a reading no longer starts with switching keyboards. The screen stopped promising more than it does: "tomorrow" on the grayed-out review button means tomorrow, practice really does not move review dates, and VoiceOver no longer speaks the word a reading question asks about.