Changelog

2026-09-16: The board wears too. Over a sitting, hairline cracks open along the grain and stains spread on the wood, each growing a little from one problem to the next and new ones appearing as the run goes on. The marks are fixed by the sitting's first problem, so they stay where they are until a break of a minute, after which the next sitting starts on a clean board.
2026-09-16: The stones are no longer one stamp. Each stone's highlight sits, stretches and shines a little differently, many have a bite out of the rim so the outline is no longer a perfect circle, a few a faint pit on the face, and some a jagged crack running in from the edge. The set is fixed by the problem, so a problem always shows the same stones with the same flaws and the next problem shows others. The stones also wear with use: a fresh sitting shows a stone or two with a defect, and the longer you solve without a break of a minute the more worn the set, until at about two hundred problems in one run nearly every stone carries one.
2026-09-16: Every motif name explains itself. Tap a line in the front-page Weaknesses list, a motif name in the Weaknesses or Tendencies tab, or an entry in the review-mode "What this tests" list, and a bubble says what the motif is and what the reading behind the line means for you, with the number in it: the rank effect on problems where that move is the answer, or the share of your wrong answers that reached for it against what the board offered. A reading that only narrowly clears the bar says so.
2026-09-16: A visual refresh of everything around the board, which is unchanged. Buttons, the menu, the dialogs and the statistics window now share one palette: a warm dark ground, off-white text, and a single vermilion accent kept for the things that mean something, namely the recall count, the rank progress and the active tab. The mode name is a small label instead of a heading, the rank progress bar has a knob and sits clear of the rank, and the ink-wash corners are quieter. The About, Changelog, Privacy, Methodology, Leaderboard and Statistics pages share the same palette and typography. The loading screen now stays up until the board is ready on a slow connection, instead of dropping onto a half-built page after fifteen seconds.
2026-09-03: Community statistics have moved to a page of their own at /statistics, which loads without the app, like the leaderboard, and can be shared by link. The menu now has a Community section holding Leaderboard, Statistics and Discord, and the personal statistics window is called My stats. Opened from the app, the page still marks where you stand in each chart; a link passed to someone else shows the community alone.
2026-09-03: The figure at the top right now says what it is. A small caption above it reads You while it shows your rank, Problem while a review shows the problem's rank, and Score once you play a move in review, where the number now stands on its own without the word in front of it.
2026-09-03: The komi circle beside the mode name is legible in light mode again. It shows what White gets, so it stays a white stone; on a light page it is now drawn the way diagrams draw one, a white disc with a dark rim.
2026-09-03: The Stats window keeps its tabs on one row that scrolls sideways instead of wrapping onto a second line, sets its text in the app's own typeface rather than a monospace one, and a tab with nothing to show yet says how many attempts it needs and how many you have.
2026-09-03: Once installed, long-pressing or right-clicking the app icon offers 9×9 and 13×13 directly, and the launch screen matches the app's dark background instead of flashing white first.
2026-09-03: The app downloads roughly a third of the script it used to on first load. Error reporting was bringing along two large modules the app never used, and a matrix library was carried for two functions that are now written into the app itself; every number they produce is unchanged.
2026-09-03: Text on the main screen — the title, the rank, the buttons and the menu — can no longer be selected. A drag or a long press that slipped off the board used to leave a highlighted word behind, and on many phones that raises a search-or-copy bar over the app, which is not how an installed app should behave. The Stats window is the exception: its tables and readings can still be selected and copied as before.
2026-08-31: The bars on the Weaknesses and Tendencies tabs now carry a scale. Thin marks sit either side of each bar’s centre line at one and two standard errors — the same two distances the colours already step at — so how far out a reading is, in the units the app judges it by, can be seen on the bar itself rather than inferred from its shade. A mark is left off where it would fall beyond the end of the bar, and the line under each table names them. The readings themselves are unchanged.
2026-08-31: Every problem has been rated afresh, by a model that is now told less about each position. Until now it was also given how strong the best moves in a position were — a fact about the answer. That is harmless for judging how hard a problem is, but the same model is what scores each individual move you might play, and there it was handing every wrong move on the board a description of the right one, which was enough to turn some readings on the Weaknesses and Tendencies tabs the wrong way round. It now works from KataGo's own read of the position and from how a range of engine settings of differing strength fare on it, and from nothing that describes the solution. Difficulty came out just as accurate — checked again against the ranks players declare on other servers, with the figures on the About page re-measured and unchanged — while the per-move readings became steadier. Separately, those readings are now weighed against a truer picture of what each problem offered. When the app works out what a typical mistake on a problem would have looked like, the available wrong moves no longer all count the same: each counts for as much as the engine actually considered it. A move it glanced at once says little about what was tempting, and a move it studied at length says a great deal, and the two were previously given equal say. Your history and your review schedule are untouched; ranks are worked out afresh from the new ratings, so yours may shift slightly as that feeds through, and the app re-reads the skill information for your recent history by itself the first time you open it.
2026-08-30: The Weaknesses and Tendencies tabs now say what their bars mean, and use their colours to say how sure the app is. A line under each table states the bar as what it is: the range the app considers plausible for that reading, one standard error either side of the number — about a 68% interval, not the 95% an unlabelled bar is usually read as. That width is deliberate, because it is also the bar a reading has to clear before the app treats it as a finding, so a row takes on colour exactly when its interval clears the centre mark. The colour now comes in two steps rather than fading in: a muted shade for a reading between one and two standard errors from neutral, the full shade beyond two, and no colour at all below one. On the Tendencies tab both directions are now the same red — reaching for a skill too often and shying away from it are the same kind of finding, and the two different colours used before invited reading one of them as good news. Which way you lean is still read off the side of the centre mark the dot sits on, and the number beside the name says it exactly. Nothing about the readings themselves, your rank or your history has changed.
2026-08-30: The skills the app tracks have been rechosen from the ground up, and there are now seven: life and death, yose, ko, opening, tenuki, sente and semeai. Every candidate skill was put through the same trial — can a player's standing on it actually be pinned down from a history somebody has, and does it stay the same skill when the model is retrained — and the seven that ship are the ones that passed both, several of them new: opening, sente and semeai join the set, and sente in particular is now defined more sharply, as the moves your opponent cannot afford to ignore where their answer to a mistake would have been to play elsewhere. Yose is the skill previously labelled counting — the new name says what it always measured. Territory leaves the list: the app could tell which problems turn on it, but not, with any useful speed, where any individual player stands on it, and a reading that never firms up is a promise the tab cannot keep. Retraining the whole set this way also made the surviving skills easier to read — a skill nobody's standing could be measured on was quietly diluting the evidence for its neighbours, so dropping it sharpened the rest. Ratings and ranks were refitted in the same pass and the accuracy figures on the About page re-measured. Your rank, your history and your review schedule are untouched; the app re-reads the skill information for your recent history by itself the first time you open it, and readings on the tab will shift as it does — those are the new dimensions being measured, not a change in your play.
2026-08-27: The Tendencies tab no longer measures your mistakes against an even split — it measures them against your own problems. Whether a wrong move "leans toward" a skill was judged with fifty-fifty as the balance point, on the reasoning that mistakes leaning neither way are balanced. But whether a mistake can lean either way is decided by the board in front of you: the answer to a tsumego is almost never a tenuki, so nearly every wrong move on offer counts as more tenuki-like than the answer — and a player picking wrong moves at random would read as leaning toward tenuki with near certainty. Measured across the whole attempt log, that one effect — the menu of moves, not the player — accounted for essentially all of the tenuki habit the tab reported, which is why nearly every player was told they had one; on other skills it was quietly pushing readings the other way. Each of your errors is now compared against what a typical mistake on that same problem would have looked like, so the number on each row is how far your own mistakes lean beyond that — and the bar is drawn the same way, balance always at the centre, the dot's distance from it exactly the number printed. Some readings will move, and a few will change sign — those are corrections, not changes in your play. With the menu no longer mistaken for a habit, tenuki also rejoins ordinary training: it was the one skill excluded from steering your problems, precisely because nearly everyone seemed to lean on it, and that reason is gone. Your rank, your history and your review schedule are untouched; the app fills in the new per-problem information for your recent history by itself the first time you open it.
2026-08-25: The privacy policy is now linked from the board itself — bottom right corner, set into the ink-wash art at the same weight as the painting, rather than being reachable only through About. It has been updated for Google sign-in in all four languages: what is stored is an anonymous identifier issued for this app alone, no address and no name, and the sign-in library is fetched from Google only if you actually open that screen.
2026-08-25: You can sign in with Google. Menu → Account has a new Google row: link a Google account once, and on any other device — a new phone, a reinstall, a browser whose data you cleared — signing in with it brings your history, rank and review schedule back without you having to remember the account code. The code still is the account and still works exactly as before; Google is a way back to it, not a replacement for it, and linking one is optional. What is kept on the server is an anonymous identifier Google issues for this app alone: no email address, no name, nothing that could identify you anywhere else. One Google account belongs to one code and one code to one Google account, and Unlink — the same row — releases both. Signing in on a device that is already using a different code asks before it swaps, naming the code you would be moving to, so a sign-in can never quietly take a device off the account it was on.
2026-08-25: Menu → Account now says what the Profile row is for. The first time you open it with nothing linked, a bubble points at that row: linking your European (EGD) or American (AGA) player record puts your name and official rank on the leaderboard, which is also the point of doing it — it gives everyone else a name to try to beat. It appears once, only while nothing is linked, and never again afterwards; the "?" beside the row explains the same thing whenever you want it. Separately, the tip that explains a repeat problem — that you have seen it before, and it has come back just as you were about to forget it — had never reached anyone. Each tip is checked when its moment arrives and again a moment later, just before it is drawn, and that second check had no way of knowing the problem in front of you was a repeat, so it refused every time. New players will now see it.
2026-08-25: You can now put your own name on the leaderboard, without an invitation. Menu → Account has a new Leaderboard section: search the European (EGD) and American (AGA) rating databases by name or member id — one box covers both, and each result is labelled with its source, official rank and country — then pick yourself. Your name and your official rating appear on the board from then on, the rating refreshed weekly from that database as before. Nothing checks that you are the player you pick, because neither database offers any way to check: what is enforced instead is that a player can only be claimed by one account, and that you can undo it. Unlink is beside the name, and takes your name and rating off the board again; linking a different player replaces the old one. The setting that keeps you anonymous has moved here too, from the foot of the leaderboard page, so both choices about how you appear now sit in one place — it is the same setting as the one in the rank survey, and works exactly as before. Invitations are gone, along with the invite tool: existing invite links no longer work, and everyone already registered keeps their name with nothing to do.
2026-08-24: The Weaknesses tab could sit empty, saying there was not enough data yet, on a device with a long history behind it — and while it did, the app also stopped aiming problems at your weak points, and the panel shown while solving named none. One stored attempt was enough to cause it. Every attempt keeps a record of what the position and your move called on, written in the terms of the skill dimensions in use the day you answered it; when those dimensions are relearned, the app drops those records and re-reads them, so they all speak the same language again. Two things could slip past that. The app updates itself between problems, but it only finds out an update exists when it reports an answer — so an app open from before an update always got to record one more attempt in the old terms first. And the problem waiting for you, and the one on your screen when the update landed, were each stored whole, with the terms they were chosen under, and were not part of what got dropped. The app now refuses to write those terms at all when it can see they are out of date, drops every place one can be kept rather than only your attempts, and re-reads anything still missing on any load rather than waiting for the next relearning. Attempts already stored in the old terms are checked once against the published data and corrected. Nothing about your rank, your history or your reviews was affected, and the two readings themselves are unchanged.
2026-08-23: The Weaknesses tab now shows every skill, always, each with two readings and a bar showing how sure the app is — in place of the short list of findings that happened to clear a bar. The first reading is whether a skill costs you accuracy beyond what the problem's difficulty already accounts for, given in ranks. The second is which way your mistakes lean, given as the share of them that would have to go the other way to be evenly balanced: if six of your ten mistakes on a skill reached for it and four shied away, one of the ten would have to change, and it reads +10%. Both are now worked out from your own attempts and nothing else. Until now each was measured against an average taken over all players — and since harder problems turn on different skills (about one problem in twenty-five at beginner level hinges on tenuki, against nearly one in three higher up), that average is largely a statement about the level you play at, so part of what you were being told about your play was your rank wearing a disguise. One skill was affected badly enough to be plainly wrong: through a flaw in how that average was applied, tenuki was reported as a fault to very nearly every player, whatever they actually did. Both numbers now stand against a fixed point of your own — no effect on your accuracy, and mistakes that lean neither way — so no one else's play enters the reckoning at all. The same two readings now also choose which problems you are served, so what the tab tells you and what you are handed to practise can no longer disagree. Tenuki is the single exception: almost every player really does lean towards it, so it is still reported to you but no longer steers your problems, since a fault shared by nearly everyone would push every player into the same corner of the collection.
2026-08-23: Over-reliance on a skill is now trained against, not just reported. When the evidence says you reach for a skill too often — playing tenuki when the position wanted something else, say — the app now serves you problems whose answer is not that skill: the situations where the temptation exists and is wrong, which is exactly where an over-used instinct needs the practice. It cannot see which wrong moves a problem offers, but it does not need to: where the answer is not tenuki, tenuki is on average among the tempting mistakes, so aiming there hits the right problems in expectation. Such a case appears in your weaknesses list as a skill you overvalue, to keep it apart from an ordinary weakness. And before claiming it, the app checks the evidence is not simply that you are good at the skill — being strong at tenuki also leaves a similar statistical trace, and that earns a strength report, not a training campaign.
2026-08-23: Your current weaknesses are now shown while you solve, and the Weaknesses tab can report strengths. While solving, the small panel that in review mode explains what a problem tests now lists — under the title "Weaknesses" — the skills your problems are currently being aimed at, strongest first. It shows the standing list only, never which skill today's problem was picked for, so it gives nothing away about the position in front of you; and when nothing is flagged it simply stays out of sight. The first time the list ever appears, a one-off bubble points it out. In the Weaknesses tab, evidence that you are unusually good at a named skill is no longer discarded: it is reported as a strength, with the same standards of evidence and the same wording tiers as a weakness — merely suspected until the evidence is firm. And the bar a skill must clear — to be listed as a problem's motif, to become a training focus, and to appear as a finding — is now one and the same everywhere, so the three can never quietly drift apart.
2026-08-23: Problems aimed at your weak points are back, chosen one weak point at a time. Once the app is confident of your rank, each skill it has reason to believe is troubling you gets a turn: a new problem is drawn for one of them at random, and it is the problem, among the fifty nearest your level, that features that skill most strongly. What it will not do is serve any single weak point more than half the time, and ordinary problems — chosen for nothing but sitting at your level — always keep a turn of their own. That restraint is not caution about annoying you — it is what keeps the diagnosis honest. If everything you saw were, say, ko problems, the app could no longer tell whether you struggle with ko or are simply having a rough patch overall: the two look identical in an unbroken run of one kind, and the misread would quietly lower your rank instead of sharpening the weakness. Mixing in ordinary problems is exactly what keeps those two stories distinguishable, so ordinary problems take a turn alongside your weak points however many of those there are: with one weak point, half and half; with two, a third each. The Weaknesses tab now also speaks with two levels of conviction: a finding the evidence merely leans towards is introduced with "We suspect:", and only one the evidence has pinned down firmly is stated outright. Which problems are chosen for a suspicion and for a firm finding is the same; only the wording differs. While the app is still working out your rank, nothing changes: it keeps picking the problems that tell it the most, as before.
2026-08-21: The black "Click to continue" screen is gone. It was there for a rare mishap — a phone short of memory can throw away the drawn board while the app is in the background — and tapping it asked for the board to be drawn afresh. Last week's attempt to keep it out of the way was to raise it a moment after the app was sent away rather than instantly; that did not help, because a backgrounded app is usually frozen outright, so the wait was never served while you were away and ran the moment you came back instead, painting the black screen straight into your return. Either way the screen was already standing there whenever the app came back, and taking it down again was a race against the first thing your phone drew — one the app lost on most switches. The app now watches for the drawing itself. A backgrounded app is not drawn at all, so the first thing drawn is proof that you are back, whether or not the phone gets round to telling the app so — and a return the app is never told about is the one case the black screen existed for. The board is rebuilt there and then, leaving nothing to tap and nothing in your way. Nothing about your rank or your history is involved.
2026-08-19: The 13×13 collection has grown by about 20,600 problems, from roughly 39,000 to roughly 68,000 — it is now well over half the size of the 9×9 one, where before it was a third. Every difficulty in the whole collection has been refitted to take them in, so your rank may shift slightly as that feeds through, and the app re-derives your stored history the first time you open it. The Weaknesses tab's skill dimensions have been relearned on the enlarged collection as well; they are the same six as before — life and death against territory, counting against shape, ko against opening — so nothing changes about what that tab can tell you. The translations of your rank into other servers' scales have been refitted at the same time. Nothing has been withdrawn and no attempt of yours has been touched. One note if you compare the About page with what it said before: the 9×9 figure there has come down slightly, from about 118,000 to about 113,000, but no problems were removed now — that number had simply been left stale since earlier withdrawals, and counting it afresh brought it into line.
2026-08-19: Opening a shared problem link could leave your next problem far too easy — a 3 dan could be handed a 2 kyu problem, and it was still waiting the following day. Your device does not hold the whole collection: it downloads it in bands of similar difficulty, keeping the ones around your own level. Opening someone else's link fetches the band that problem belongs to and nothing else, so for the rest of that visit the only problems on your device were the ones near the difficulty of the problem you were sent. The app picks your next problem as soon as you answer the one you are on — and during such a visit it had nothing to pick from but that band, so it took the closest match it could find there, which could be a long way from your level. That choice is remembered for next time, which is why it survived until the following day. The app now chooses only from an unbroken stretch of downloaded bands that takes in your own, so whatever it picks is genuinely near your level even when little has been downloaded; and if your own band is not on the device, it fetches that first rather than settling for what happens to be at hand. It keeps the bands just outside the ones it is choosing from ready in advance, so none of this costs you any waiting as your rank moves. When your band truly cannot be fetched — with no connection, say — it falls back to the nearest bands you do hold and tells you it is serving away from your level, exactly as before. Nothing about your rank or your history was affected, and shared links themselves are unchanged: they are still view-only and still never counted.
2026-08-14: Leaving the app and coming back could put a black screen reading "Click to continue" in your way — on Android often on every single switch. That screen is there for a rare mishap: a phone short of memory can throw away the drawn board while the app is in the background, and tapping the screen asks for it to be drawn afresh. It was being raised the instant the app was sent away, but a phone carries on drawing an app for a moment after that, while it slides off the screen — so the black screen was caught in the picture the task switcher shows you, and was still standing there when you came back. It then outstayed its welcome, because the app rebuilt the board before allowing the screen to be taken down, and the black screen was the last thing drawn until that finished. It is now raised only once the app has stopped being drawn, so an ordinary switch never catches it, and on your return it comes down first and the board is rebuilt after. It still appears, and can still be tapped, on a return the app is never told about — which is the case it exists for.
2026-08-13: For a few players the wooden board texture had vanished, leaving a flat tan board behind — and once it happened it stayed that way, through reloads and through every update. The app keeps its own copy of the handful of files it always needs, so that they load instantly and still work with no connection. Through a fault in how those copies were stored a second copy of the board texture could be left behind, and from the first update onwards the app read that one instead of the fresh one; if it had been saved badly there was nothing to put it right, because a background picture is only ever asked for once. Those files are now always read from the single place that is rebuilt on every update, any stray duplicate is cleared out, and a hiccup while saving one of them no longer costs the whole set. Affected boards come back on the next update. Nothing about the problems, your history or your rank was involved.
2026-08-13: Review reminders no longer make a sound, and they now count down as you work through them. A reminder has nothing urgent attached to it — the problems are due, not about to vanish — so it arrives quietly and waits in the notification tray rather than announcing itself. Once it is there it keeps up with you: each review you answer drops the number it shows, and clearing the last one takes the reminder away by itself. Until now it was meant to disappear as soon as you touched the board, on the reasoning that by then it had done its job — but on Android it usually did not disappear at all. A reminder delivered while the app is closed can no longer be reached by the app once you reopen it, so it sat in the tray indefinitely, still claiming problems were waiting long after they had been reviewed. Nothing changes about which problems are due, or when a reminder is sent.
2026-08-10: Once the app is confident of your rank it starts choosing problems by the skills that have been tripping you up — and that choice was barely working. Two things were wrong with it. First, it judged how well a problem matched your weak points on an absolute scale, and almost all of what it was reading turned out to be identical for every player: a standing preference for certain kinds of position, coming from the way wrong answers tend to differ from right ones across the whole collection rather than from anything about you. Everyone was being steered towards the same corner of the collection and told it was personal. What every player's reading has in common is now measured and taken out, so what remains is only how you differ from everybody else. Second, having worked out which candidates suited you best, the app then picked among them close to at random — the differences between problems at your level are small, and the way it weighed them left the choice almost indistinguishable from ignoring the weakness reading altogether. It now simply takes the best candidate, both while it is still working out your rank (where it takes the problem that tells it the most) and afterwards (where it takes the one that best fits your weak points). Together these make the weakness-driven choice about eight times as strong as before, and it now points somewhere genuinely different for different players. You will not see less variety: a problem you have answered is not offered again for a long time, so the best remaining candidate is a different problem each time.
2026-08-09: The About page now says what the app's accuracy is worth in games. "Within one rank of what a player declares" is hard to picture, so the same comparison is also stated as a tournament record: how many games an ordinary rating would need in order to be this accurate. Based on published statistics from the European Go Database, and measured on the players whose European rating is machine-verified rather than self-reported, the rank shown here agrees with a player's real rank about as closely as two independent ratings built from about sixteen tournament games agree with each other, give or take three. Nothing about your rank changes; this is a new way of describing how good it already is.
2026-08-09: The whole collection has been re-checked, with a stricter test than the one used a week ago — and that test says the earlier one was wrong in both directions. 1,374 problems withdrawn on the 2nd and the 8th are back: they were sound all along, and the checks that removed them had been reading lines the computer started exploring and then abandoned, treating an unfinished line as a verdict. In the other direction, 4,032 problems have now been withdrawn that those checks missed. The new test only looks at lines played all the way to the end — both players passing, the game genuinely over — and asks two things of each problem: that the move you are asked to find still wins when the opponent tries everything recorded against it, and that every other first move really does lose. Anything short of the end is ignored rather than guessed at, which is exactly what went wrong before. Your attempts on withdrawn problems are kept this time. Last week's withdrawal took your history with it; this one does not, so your totals and solve counts stay as they were, and only problems you have not yet seen are affected. Difficulties are being refitted on the corrected collection, so your rank may shift slightly as that feeds through.
2026-08-09: Stepping back through an answer now follows the lines that were actually played out, rather than working out which to show each time you open one. Each problem file now says outright which continuations were read to a finish, and the app follows those. Before, it looked for any pass anywhere below a move and took that as a sign the line had been seen through to the end — which is not the same thing, and is how a continuation that had never been properly explored could be offered to you and then drift into a lost position. The opponent's replies you are shown are chosen exactly as before, so you will still sometimes see two of them; they are simply now drawn from lines that genuinely finish.
2026-08-08: Just under 1% of the collection has been withdrawn — 1,124 problems where stepping through the answer could walk you into a lost position. The first move of every one of them was right, and winning; the fault was in what came after it. When you look back over an answer the app offers the opponent's likely replies, and for these problems one of those replies opened a continuation that had never been played out properly — follow it and the answering side drifts from winning to lost, ending on a losing score, even though nothing was wrong with the problem or with the move you were asked to find. Checking whether a line reaches the end of the game is not the same as checking that it was played well, and these had only ever been checked the first way. The problems are gone, and any attempts you had on them have gone with them, so your totals and history will show slightly fewer solves than before. It was about twice as common on 13×13 as on 9×9. Difficulties have been refitted on what remains and the Weaknesses tab's skill dimensions relearned from them: reading against shape no longer matched how players actually differ and has been retired — as always it is kept on file and comes back by itself if it reappears in future data. Your rank may shift slightly as the refreshed difficulties feed through, and the app re-derives your stored history the first time you open it.
2026-08-08: You can now call up a review yourself. While you are solving, the button that used to sit greyed out as Share is a Recall button, and it brings up the problem that is most due to be seen again — with a small badge showing how many are waiting. Reviews still arrive on their own, spaced out so they do not crowd out new problems, but until now that was the only way to reach one: if several were waiting and you wanted to clear them, there was nothing to press. The button is greyed out when nothing is due and while you are already on a review. Nothing about the schedule changes — which problem is due, and when, is worked out exactly as before, and the new problem you were on is not lost, it comes back as soon as you have answered the review. Share is unchanged, and still takes that place while you are reviewing an answer.
2026-08-06: The screen after an answer no longer congratulates you on reaching a rank you already had. How cautious the app is about the rank it shows depends on how sharply the problems around your level tell one player from another, and it worked that out from whichever problems your device had downloaded at that moment — so when it had none yet, at the start of a session or just after restoring an account on a new device, it had to assume the worst, and your rank was shown as much as a whole rank low. The next problem you answered put it back, and the app read that as a promotion and congratulated you on a rank you had held all along. That judgement is now measured once across every published problem and shipped with the app, so it is the same before you have loaded anything as after, and a promotion means your play moved. Crossing back and forth over a rank boundary still reports each crossing, as it should — that one is real.
2026-08-06: Your rank is now worked out the same way everywhere. The same judgement above was made separately by your device and by the server that builds the leaderboard, each from a different set of problems, so the rank on your front page could disagree with the one on the board — and it could shift slightly as you played, purely because a different set of problems had loaded. It is now measured once and shipped with the app, so your badge, the leaderboard and the Community chart all read from the same figures. You may see a one-off correction of a fraction of a rank, and new players start a little closer to the middle of the scale than before.
2026-08-06: The leaderboard can now be read in any of the ranking systems the app knows — OGS, KGS, Fox, Tygem, WBaduk, IGS, GoQuest, EGF and AGA, as well as onlymove's own — and it now shows what players say their rank is, not only what the app predicts. The two columns are named for what they are: Declared is the rank the player themselves reported on the selected system, and Predicted is the app's estimate of it, so the two can be read side by side in whichever system you play on. For a player who has linked their EGD or AGA account, the verified rating from that account is shown instead of their own report in that system's units. The rank survey now says what your answers are used for and carries a checkbox to keep your declared ranks off the board — the same setting as the one at the foot of the leaderboard itself, so ticking either hides them, and your predicted rank still appears either way, anonymously. The page is now titled Weekly Leaderboard, since it has always listed the past seven days' active players, and the chosen ranking system is carried in the address, so /leaderboard?units=Fox opens on Fox and a link you copy shares the view you are looking at.
2026-08-06: The Community tab's rank chart now marks you at your own rank. The bars are each player's actual rank, but the marker showing where you fall was still worked out separately, from the difficulty of the problems you happened to be served that week — so it could sit a whole rank away from the rank on your own front page. It reads high while the app is still unsure of you, since it deliberately serves a little above your level then; and once it starts choosing problems by which skills have been tripping you up, what you are served stops tracking your rank at all. Your marker is now simply your rank, binned exactly as everyone else's bar is, and it appears only once you have solved enough problems to carry a rank — the same threshold the bars use. Your rank itself is unchanged.
2026-08-06: On the leaderboard, each value now reads as the rank followed by its progress percentage — "4d (39%)" rather than "(39%) 4d" — in every column and every unit setting, so the rank itself comes first.
2026-08-06: Tapping one of the tip bubbles that greets a new player no longer plays a move. Three of those bubbles sit over the board, and although they look like something you can tap away, the tap went straight through to the point underneath — on a computer, where a single click plays a move, that answered the problem there and then, almost always wrongly, before the tip had even been read. Tapping the bubble now only dismisses it. Tapping anywhere else still dismisses it and does whatever you tapped, so nothing else about them changes. This only ever affected players new enough to be shown these bubbles, in their first few problems.
2026-08-06: The screen shown after an answer no longer reports a change that disagrees with the answer. Both numbers it can show — the rank movement, and the change in repeat delay after a review problem — come from estimates the app is still refining, and either could occasionally land on the wrong side of zero: a small rank gain after a wrong answer, or a review pushed further away after failing it. Either now reads as the smallest movement the line can show, in the direction your answer earned, rather than contradicting the result you just got — a correct answer is never reported as having bought you nothing, and a wrong one never as a gain. A rank line is shown after every answer instead of being left out when it happened to be zero. This is only what the screen says: your rank and your review schedule are unchanged, and still exactly what the app worked out.
2026-08-05: The rank scale down the side of the Rank tab's chart was labelled half a rank too low, so the graph read a whole rank stronger than it should — a player at high 3 dan saw their line sitting in a stripe marked 4 dan. The labels were placed for an older way of writing ranks, where a whole rank was centred on its number; the app now treats those numbers as the boundaries between ranks, and the chart had not been moved with it. The gridlines now fall on the boundaries and each rank is named in the middle of its own stripe, so the chart agrees with the rank stated underneath it and with the front page. Only the chart's scale was affected — no rank, and nothing the app did with it, was ever wrong.
2026-08-05: Your very first problem is now chosen by the same rule as every later one, and solving it moves your rank straight away. A brand-new account was handed a problem at exactly the rank it was showing you — the one case where the app aimed at what it was displaying rather than slightly above it — so the first correct answer moved the estimate without ever moving the number on screen, which looked like the answer had been ignored. New accounts now start on something a little harder, and while you have not missed anything yet the app reaches a couple of ranks further still. A first correct answer now moves the shown rank by a full rank. The extra reach ends the moment you miss a problem, so it only ever applies while your record is clean.
2026-08-05: Anyone much stronger than the rank they start at now reaches their real rank in roughly half as many problems. The app deliberately serves a little harder than the rank it is showing you, so that it keeps learning where you actually are — but how much harder was a fixed five ranks, regardless of how much or how little it knew about you. That reach is now earned: one rank for every problem you clear in a row, reset whenever you miss one. The first problems are gentler than before as a result — a brand-new account is offered something at its own level rather than five ranks above it — while a run of correct answers lets the app climb to meet you much faster. A strong player starting from scratch now reaches a dan rank in around twenty-five problems rather than forty. Your rank itself, and the deliberately cautious way it is shown while still uncertain, are unchanged.
2026-08-05: Added a methodology page — the full account of how a problem's difficulty is measured, how those difficulties become the rank you are shown, and how the weakness dimensions are found and used. It is a technical read rather than a summary, written for anyone who wants the reasoning behind the numbers. It is the project's own working document, published straight from the source with each release, so it cannot drift out of date with the app. English only; linked from the About page.
2026-08-04: The bubble that greets a new player on their very first problem now describes how moves are actually played on a touchscreen: tap the point, then tap it again to confirm. On a phone or tablet the first tap only places an outlined stone — a second tap on the same point plays it, and a tap elsewhere moves it — but the bubble described only the first tap, so a first-time player could be left tapping once and wondering why nothing happened. On a computer, where a single click plays the move, the wording is unchanged.
2026-08-04: The rank on the front page is worked out a new way, aimed at the complaint that it used to jump around and read too high in your first sessions. It now starts from a measured picture of how strong onlymove's players actually are and updates that with your answers, instead of starting from wherever your own estimate happened to have drifted to — which is what let an early run of right or wrong answers carry itself forward. While it is still unsure of you it deliberately shows the cautious end of its estimate rather than the middle, so the number climbs towards your real rank as evidence comes in instead of overshooting and having to walk back down; once it is confident, the caution disappears and the shown rank settles on the estimate itself. Problem selection follows the same number: problems are chosen a little harder than what is shown, so the app keeps learning about you, but never more than five ranks above it — you will not be handed dan problems while it is calling you a beginner. The Rank tab's front-page-rank line now traces exactly what you were shown, on OnlyMove and on every server option; the steadier skill line beside it is unchanged and still shows the underlying estimate.
2026-08-02: Just under 3% of the collection has been withdrawn — 4,708 problems whose answer did not survive being played out. Every problem is checked at creation by playing both the winning reply and the best losing reply through to the end of the game, but the move-by-move evaluations stored in the problem file come from separate, shallower readings, and a sweep of the whole collection found places where the two disagree: the winning line finishing in a loss, or the refutation of a wrong move finishing in a win. Those problems are gone, and any attempts you had on them have gone with them, so your totals and history will show slightly fewer solves than before. The disagreement was far more common on 13×13 than on 9×9, since those lines run longer and turn on ko much more often — exactly where a quick reading and a deep one part company. A further 348 problems kept their place but had their main line completed: it stopped short of the finish, usually one pass before the game was actually over, so stepping through the answer ran out of moves early. Difficulties and the Weaknesses tab's skill dimensions were refitted on what remains, so your rank may shift slightly as that feeds through.
2026-07-28: Every problem's difficulty has been refitted against the latest solving record, and the Weaknesses tab's skill dimensions were relearned from it. A new dimension separates killing — positions whose answer is the move that takes an opponent group off the board — from ko threats, positions that turn on finding a threat elsewhere. Two dimensions that shipped previously, attack/defence and reading/shape, no longer matched how players differ once the difficulties were refreshed and have been retired; as always they are kept on file and return by themselves if they reappear in future data. Separately, the app's internal rank scale was rebuilt onto a single consistent footing — every rank you see reads exactly as before, but the app rebuilds its locally stored data and re-learns your personal review and weakness models the first time you open it after this update. Your rank may shift slightly as the refreshed difficulties feed through.
2026-07-28: Once you get a review right, the next sight of that problem now comes about twice as far apart. Checked against the solving record, reviews were coming back too soon after a first success — they were being cleared around 80% of the time and above, where the schedule aims for nearer 70%, meaning they arrived before you had really begun to forget. The delay after a missed problem was already right and is unchanged. The cause was that the app treated "you just missed this" and "you just got this right" as two steps of one ladder rather than the different situations they are: getting a missed problem right again restores more than an ordinary success's worth of memory, and that is now accounted for.
2026-07-27: Reviews of failed problems now come round somewhat more often. The pace is still set by how much you play and how quickly you personally forget, but it is now matched to a lighter day than your recent average rather than to the average itself — so a spell of playing less than usual no longer leaves due reviews piling up, and at your normal pace you get through them with room to spare.
2026-07-27: Every problem's difficulty has been refitted against the latest solving record, and the Weaknesses tab gained a new skill dimension separating attack from defence — one side covers positions whose answer is to press an advantage, the other positions whose answer is to save what you have. It takes the place of a dimension that no longer matched how players actually differ; the old one is kept on file and returns by itself if it reappears in future data. Your rank may shift slightly as the refreshed difficulties feed through, and the app re-derives your stored history on first load.
2026-07-27: The Community tab's rank distribution is now built from each player's actual rank, the same estimate the front page shows them and the same one the leaderboard ranks them by. It used to be inferred from the difficulty of the problems people had been served that week, which is only approximately their level: the app deliberately serves slightly above your rank while it is still unsure of it, so newer players read too strong, and the difficulty was rounded to the band of a whole problem set, which reads strong players too weak. The chart also now counts the players who carry a rank — the same 30-problem threshold the leaderboard uses — rather than anyone who answered a handful of problems in the week, so it no longer mixes in players whose rank isn't yet established. The country list on the same tab now shows every country instead of the top fifteen and an "Other" row.
2026-07-26: The Weaknesses tab now lists fewer, more trustworthy issues. Each candidate issue used to be judged on its own against a fixed confidence bar, so a borderline one was listed whenever it cleared that bar — and since every skill dimension is checked for every player, and the app gains dimensions as the problem collection grows, the list would slowly fill with entries that were right only about four times in five. The listed issues now share one error budget: they are added strongest-first for as long as the combined chance that any of them is spurious stays small. A single strong finding is listed exactly as before; a long run of marginal ones is now cut short.
2026-07-26: Problem selection and rank convergence now account for how guessable each problem is. A problem you can stumble into by luck tells the app less about your strength than a sharp one you clear just as often, and that was previously ignored in two places: guessable problems were slightly over-weighted when choosing your next problem, and the app's threshold for "your rank is now pinned down" assumed a precision no amount of solving could actually reach. The second one had a visible effect — problems kept being served a fraction of a rank above your level indefinitely, instead of settling onto your own rank, and the switch to weakness-targeted problems came later than intended. Both now use each problem's full response curve.
2026-07-23: Moves can now be played from a physical keyboard, as an accessibility feature for anyone who finds precise tapping or mouse work difficult: type a move's board coordinates — e.g. C3 — anywhere in the app and press Enter to play it (columns skip I, rows count from the bottom; works while solving and while reviewing). To make coordinates readable, tiny coordinate labels now line the very edge of the board, small enough to stay clear of stones on the first line. Nothing changes visually for anyone else — the entry field itself is invisible.
2026-07-23: Offline play now lasts much longer: when the problems near your rank run out and the next problem set can't be downloaded, the app now serves problems from the sets already stored on your device — further from your usual level — instead of silently showing the same problem again after each answer. A brief notice appears when this fallback starts, and another when normal at-your-level serving resumes; only if nothing stored remains unsolved does the app re-serve your last problem, now with a clear explanation instead of looking like a bug. Answers to such a re-served problem were already excluded from your rank and stats.
2026-07-22: Play Store (Android) app: pressing Notify now brings up Android's own notification permission dialog directly (app version 3). The web app asks the app to request the permission itself instead of relying on Chrome's broken relay, so the manual settings walkthrough from earlier today is only shown on older app versions that can't do this yet — and if Android would silently refuse the dialog (permission previously denied, or notifications turned off in settings), the app opens the right settings screen for you instead.
2026-07-22: The result overlay after an answer now appears immediately. Previously two things could delay it by up to a few seconds: after a review (repeat) problem, the retraining that computes the "repeat delay" line ran before the overlay (it now runs behind it, and the line fills in as soon as it's ready); and after solving a new problem, picking your next challenge — sometimes including a problem-set download — also ran before the overlay (it now happens under it).
2026-07-22: In the Play Store (Android) version, enabling review reminders now also walks you through allowing notifications for the app in Android's settings. Android 13+ blocks an app's notifications until you allow them, and the system dialog that should ask on the Notify press doesn't appear (a known Android/Chrome issue) — so reminders were silently dropped and the app "forgot" the Notify setting within a day. Until that's fixed upstream, the app now shows the steps to allow notifications right after you press Notify; with the toggle on, reminders both arrive and stay enabled.
2026-07-22: New players who miss three of their first ten problems now get a small tip bubble: the correct answer is usually the natural response to the opponent's last move — an encouragement not to over-analyze the early problems.
2026-07-21: The leaderboard got a units selector — view the board in onlymove, EGF, or AGA units. The former Declared column is now Verified: it shows only the official rating of a player's linked EGD or AGA account, refreshed weekly from those databases — self-declared survey ranks are no longer shown on the board. A rating appears only while it is active by the EGF's activity criteria: a tournament within 2 years for dan players, 1 year for single-digit kyu, 6 months for double-digit kyu. Every value reads as the rank with progress toward the next — e.g. a 2389 GoR shows as (39%) 4d, the same style as the front-page rank list — and with EGF or AGA units selected, the onlymove estimate is converted into that federation's own scale for a direct side-by-side comparison; in onlymove units the Verified column stays empty. On phones the page now opens zoomed out to fit the screen width exactly.
2026-07-20: You can now link an American Go Association (AGA) rating to your account, the same way as a European (EGD) rating — through an invite link. The invite tool now has a single search that queries both the EGD and AGA databases at once and lists the results together, each labelled by source; pick a player and their current rank is taken straight from that database (refreshed weekly). A linked AGA player then appears in the leaderboard's Declared column on the onlymove rank scale, alongside EGD players. Each account links to a single rating identity — EGD or AGA.
2026-07-20: The leaderboard page, the About and Privacy pages, and push review reminders are now shown in your language — Chinese, Japanese, and Korean, alongside English — matching the rest of the app. This changelog moved to its own page (the one you're reading); the About page links here.
2026-07-20: The leaderboard now lists every player active in the past week who has solved at least 30 problems, applying one threshold to everyone instead of treating invited and uninvited players differently. A short line at the foot of the page counts how many active players haven't reached 30 solves yet, in place of the per-player progress that used to show for not-yet-ranked players.
2026-07-20: You can now hide your profile from the leaderboard — a control at the bottom of the page withholds your name, country, and declared ranks (your onlymove rank still appears, but anonymously), and can be switched off again at any time. The leaderboard also now shows a country flag for every active player: their EGD federation or rank-survey country where known (preferring the more reliable EGD one), otherwise the country their play connects from, and a neutral flag when even that isn't available — so every row, anonymous players included, carries a flag. The Community stats tab now also shows a breakdown of active players by country. And both the onlymove estimate and the Declared range now read on the same onlymove rank scale, without the solving-speed adjustment, so they line up with your front-page rank. The Finnish tournament has concluded, and its page now shows the final standings.
2026-07-19: The leaderboard now lists every player active in the past week, not just invite-registered players, and speaks ranks instead of rating points. Unregistered players appear as Anonymous with their estimated rank — computed from their last 500 first-time problems, the same way as the front-page rank — while registered players show their name. A new Declared column shows where each player's external ranks (their official EGD rating and any server ranks declared in the rank survey) land on the onlymove scale — shown as a range when the sources disagree — next to the onlymove estimate for comparison. Registration remains invite-based. The this-week/last-week peak views were removed — the board (and the tournament standings) always show current estimates.
2026-07-18: The weakness system now runs on skill axes learned directly from how players of the same rank differ in what they actually solve — replacing the previous motif clusters derived from KataGo's internals alone. Each axis is validated to predict future results before it ships, and axes are two-sided: for example, one axis separates early open positions from late settled ones, and another separates positions that punish shape intuition from lines that reward it. New axes are added automatically as the data grows. Problem difficulty discrimination was also recalibrated (response curves were measurably steeper than modeled), improving how quickly your rank estimate converges.
2026-07-17: In the Play Store (Android) version of the app, the Notify button now correctly turns into the inactive "Reminders" state after you allow notifications, instead of staying clickable — reminders themselves were already being delivered; only the button state was wrong. Recall-schedule updates from the Play Store app are now also uploaded reliably during play.
2026-07-17: Weakness example problems now open for review inside the app itself instead of in a new tab — the new tab collided with the one-tab-at-a-time guard and could show "already open in another tab". The stats panel steps aside while you review the example and returns exactly where you left it when you tap Next, and the Previous button still takes you to your genuinely previous problem afterwards.
2026-07-17: The detailed stats now open much faster on accounts with a long play history — the Rank tab's chart was the slow part, and building its long-term skill curve and bands is now far cheaper, so opening stats no longer stalls for many seconds on large accounts.
2026-07-17: Only Move is now on Google Play. On Android, when your browser can't install the app directly (Firefox, Samsung Internet, and others), the Install button now offers the Play Store version — the same app — instead of only showing manual instructions. If you already have the Play Store app installed, the browser no longer prompts you to install again.
2026-07-16: Reviews of failed problems are now interleaved with new problems at a pace matched to how much you play and how quickly you personally forget, instead of every due review always cutting ahead of new problems — so heavy review days and walls of reviews after a break are gone, and the balance settles to what your own volume and memory sustain. Well-learned problems now also genuinely retire: each success stretches the next review out further without limit, where previously even a fully mastered problem kept coming back every three months or so forever. Review reminders from notifications are unaffected and always open the due problem.
2026-07-14: Displayed ranks are now rounded to the nearest whole rank instead of always rounded down, matching the standard European (EGF) rank convention — so your shown rank may go up by one (a strength just past the three-/four-dan boundary now reads 4 dan rather than 3 dan). This is purely a display correction: your underlying strength, the problems you're served, and your estimated per-server ranks are unchanged. Tapping your front-page rank now lists OnlyMove alongside the other servers as one list, each showing how far you are into that rank, and the 95% confidence interval on your rank moved to the stats Rank tab. On the Rank tab, the solving-speed adjustment for the timed server views now uses the same recent window of problems as your rank, keeping the two in step.
2026-07-11: The Rank tab's chart now also plots your front-page rank over time as a solid line, alongside the smoothed long-term curve — now named "Stable rank" — so recent form and settled strength read off one chart. Rank calibration is now robust against the occasional clearly mistaken survey declaration.
2026-07-10: Removed about 500 problems that had more than one working solution, and the generator now rejects such positions outright; problem ratings were refreshed across both board sizes after the cleanup. When a ratings refresh like this retroactively shifts your rank history, the app now explains it with a brief one-time notice. The analysis hint now always marks the genuinely strongest move. Weakness-aware problem selection now uses the direction of each weakness: motifs you under-play come up more often, while motifs you over-play are no longer served as if you needed more of them.
2026-07-09: Cross-device sync matching is now exact: your device stamps each attempt's time itself, so an attempt can no longer be duplicated or missed when syncing. The stats summary now splits your solved-problem count by board size. The manual install instructions now match your actual browser and platform.
2026-07-08: Reworked the after-answer layout: the Share button is now always visible, and the "why this move" motif bubble aligns consistently with the buttons in every orientation. The rank survey got a proper centered rank picker on desktop. Rank calibration now anchors each survey answer to your strength around the time you gave it rather than your whole career, which reads improving players more accurately. Account sync was rebuilt on exact server-record identity, fixing rare accounts that could re-download the same history over and over.
2026-07-07: The share image can now be copied straight to the clipboard on desktop. Wrong moves are now recorded independently of the orientation the board was shown in, so they sync exactly across devices. Tuned caching so repeat visits load faster.
2026-07-06: The Community "how quickly players return to failed problems" chart now measures the gap from first failing a problem to its first review. Opening the app in two tabs at once can no longer record duplicate attempts. Syncing your history onto a new device now downloads a small fraction of the data it used to. A failed account adoption now falls back cleanly to a fresh account instead of getting stuck halfway.
2026-07-05: Returning to the app now syncs immediately and refreshes the problem on screen — a problem you solved on another device is swapped out behind a brief notice, and a review that came due while you were away can appear right away.
2026-07-04: Problem difficulty targeting now converges correctly at every rank — kyu players were previously served problems slightly above their level on average. Added a landing page for the Finnish problem-solving tournament (July 13–19) at onlymove.net/tournament.
2026-07-03: Tapping your front-page rank now shows a 95% confidence interval on your rank, so you can see how precisely it's currently pinned down alongside your estimated rank on each Go server. Refreshed problem ratings across both board sizes and the weakness-motif labels.
2026-06-30: Reworked how your rank is calibrated, correcting a long-standing under-rating of stronger players — so some dan-level ranks will rise. The solving-speed handling is also clearer: your front-page rank is not adjusted for speed, because here, thinking longer genuinely makes you more likely to solve a problem, so that rank reflects your strength with all the time you want. Your rank on each individual Go server is adjusted, because in a real game a slower time control doesn't make you more likely to win — your opponent gets the same extra time and is just as much stronger for it — so leaning on long thinking reads a little weaker on a timed server.
2026-06-29: Added about 22,000 new 13×13 problems, roughly doubling the 13×13 collection, and refreshed problem ratings across both board sizes. Fixed a bug where the stats and rank graphs — and the problem ratings behind them — could differ between two devices signed in to the same account.
2026-06-28: The "current rank" end of the Rank tab's skill curve no longer lags behind your recent improvement, so the most recent point of the curve reads your present strength more faithfully.
2026-06-26: The front-page rank is now a maximum-likelihood estimate fitted over your recent solving history — the same kind of computation as the Rank tab's long-term curve — replacing the older momentum-based rank. It reads your recent form more cleanly and no longer depends on the order you happened to solve problems in.
2026-06-25: The problem you're solving no longer flips to a different orientation when the app quietly updates itself in the background — the board stays exactly as shown across a reload.
2026-06-24: The Community "Attempts per day" chart now averages each player's attempts over the full week (weekly attempts ÷ 7) for a true daily average, instead of counting only the days they were active, and its ranges were rescaled to match.
2026-06-22: Rebuilt the problem-rating model again. It now learns its overall difficulty scale, easy-guess floor, and reliability ceiling automatically instead of having them set by hand, and it reads the strongest candidate moves in each position as an extra signal. Problem ratings have been refreshed across both board sizes, and the weakness-motif labels were refreshed to match the rebuilt model.
2026-06-21: Finer-grained bins on two Community charts — "how quickly players return to failed problems" and problems solved per day — so the busiest ranges read more clearly.
2026-06-20: Review reminders now reach every device signed in to your account, not just the most recently used one. Fixed a recall-scheduling bug so failed problems reliably come back for review on the right board.
2026-06-19: The detailed stats page now opens at any attempt count, instead of waiting until you had solved several problems. The Community time charts switched to a clearer log-scaled layout with empty buckets dropped. Sharing now starts preparing the image the moment you open the share chooser, so it feels snappier.
2026-06-18: Multiple pop-up messages now stack instead of overlapping. Review mode survives closing and reopening the app — the problem you were reviewing is restored. Made sharing more reliable after the app has been idle or when running in a less-secure context. Long solving histories are now stored more durably for heavy users, so they no longer risk hitting the browser's storage limit. Fixed a bug that could occasionally serve the same new problem twice in a row.
2026-06-17: Fixed a bug where reloading right after solving could re-serve the problem you had just completed.
2026-06-16: Corrected a handful of problems whose stored solution did not match KataGo's best move — the genuinely winning move is now always marked correct.
2026-06-14: Refined the time-adjusted rank so your solving speed is compared fairly against players of a similar rank. Problems opened from shared links are now fully excluded from the server logs as well as your own stats. Made problem downloads more resilient to flaky network connections.
2026-06-13: Your progress is now a portable account. Open Account (menu → More) to see your personal code, then enter it on another device to sync your full solving history and rank across all your devices. The Rank tab's chart also gained a server selector — switch between OnlyMove's own rank and how your strength would read on each Go server, adjusted for your typical solving speed. Tapping your front-page rank shows the same per-server, time-adjusted ranks.
2026-06-12: Added a sound on/off toggle to the menu. Tapping a review reminder now always opens whichever problem is actually due, rather than one you may have already cleared. Corrected several details in the stats and rank math for more accurate recall difficulty, weakness trends, time comparisons, and confidence bands.
2026-06-07: Added a link to the community Discord, both in the menu and at the bottom of this page. A rare misclick on an easy problem no longer dents your rank. Problem difficulty is now centered so you clear about 80% of the problems you are served, and some front-page ranks that had drifted were re-anchored.
2026-06-06: Added five new weakness motifs (with Chinese, Japanese, and Korean translations) to the recommender and stats page. The "comes back in about X hours" estimate shown when you fail a problem is now based on your own recent recall performance rather than idealized assumptions. The Time tab now shows how confident the app is that thinking longer actually helps you. Added a brief loading splash so the app no longer flashes a half-built screen on startup.
2026-06-05: The post-failure "don't rush" and "review more" warnings are now bias-corrected and only appear when there is real evidence that slowing down would help you. Improved the estimate of your chance of solving an unfamiliar problem, including the realistic chance of a lucky guess on hard ones.
2026-06-04: Added decorative ink-wash corner artwork that adapts to light and dark themes. Strong players can now climb past 9 dan, with rank gains smoothly slowing at the top. The skill estimate on the Rank tab now focuses on problems near your own rank for a cleaner read of your progress.
2026-06-02: The screen after each attempt now shows how your solving time compared with your usual pace. Refined the Time chart with whole-second and whole-minute axes, date gridlines, and a middle-50% spread band.
2026-06-01: Polished the Time chart with a smoother trend line and a cleaner spread band, and your own bucket is now highlighted in each Community chart. Stats charts now build behind a "Loading your stats…" overlay so they appear all at once.
2026-05-31: Rebuilt the problem-rating system as a single model shared across both board sizes, tuned to keep its accuracy even across all difficulty levels. Problem ratings have been refreshed accordingly.
2026-05-29: Added a Community graph of how quickly players return to failed problems. Switching board size from the menu now jumps straight to the other board. Reworked the spaced-repetition schedule to an FSRS-style retention curve, and made the app download smaller and faster.
2026-05-28: Added a Community tab with server-wide statistics from the past 7 days — overall activity, rank distribution, time usage, and recall times across all players. Grouped the menu into Practice / Review / More, and added a daily solving- and review-time chart to your stats.
2026-05-27: Moved About, Stats, Share, Ignore, and the board-size toggle into a hamburger menu. Added a Pause recalls / Resume recalls toggle to temporarily stop failed problems from reappearing.
2026-05-22: 9×9 and 13×13 now share a single rank, stats page, and learned models while keeping separate problem histories. Your stats now show how your 13×13 skill compares with your 9×9 skill.
2026-05-21: Added a review-mode bubble naming the motif behind the correct move. The skill chart now shows a continuous skill curve with separate measurement and performance bands.
2026-05-20: Your rank is now anchored to European Go Database ranks and translated into the rank system of whichever server you play on.
2026-05-18: Added an occasional, optional survey asking your rank on other Go servers, which helps calibrate the ranking for everyone.
2026-05-15: Per-attempt details (the move you played, solving time, time spent in review, and whether the problem was a repeat) are now sent to the server alongside the pass/fail result. The weakness section of the stats page is now a single sorted list of plain-English sentences ("You don't ___ often enough.", "You ___ when you shouldn't.", "You inconsistently ___.") instead of three separate panels. The app now auto-reloads between problems whenever the server has a newer build. Refreshed embedding labels with a few typo fixes and better-distinguished motifs.
2026-05-14: New rank chart on the stats page with a smoothed long-term skill estimate, a 95% confidence band, and a session-to-session form variance metric. Added a per-attempt scatter chart showing every result against your rank over time. Push notification reliability fixes: empty schedules no longer accidentally wipe pending reminders, notification taps reliably focus a backgrounded PWA on Android, and the recall reminder shows total backlog rather than just the delta since the last reminder. Tightened the math behind the weakness quadrant probabilities so anti-correlated estimates no longer show false "both directions" mass.
2026-05-12: Added support for 13×13 boards — visit onlymove.net/13 for the larger boards (9×9 stays at onlymove.net/9). Added a weakness-aware problem recommender: every position is mapped to 64 motif axes derived from KataGo's internal representation, and a per-user logistic regression tilts upcoming problems toward motifs that have been tripping you up. On wrong attempts the move you actually played is mapped into the same space, so the model can tell apart "the position is hard to read" from "you over- or under-play this kind of move".
2026-04-25: Improved sharing: no more lag, immediate feedback while preparing an image, clear error messages on failure, and tapping the share button repeatedly no longer opens multiple dialogs. Fixed the inflated "ignored problems" count in stats — it now only reflects problems you explicitly chose to ignore. The "don't rush" and "insufficient review" messages now compare against your median solving and review times across recent problems, for more responsive feedback. Fixed the ghost stone in review mode disappearing after the app was backgrounded and resumed.
2026-02-13: Removed 6k problems where the mainline didn't stay consistently ahead for black.
2026-02-07: Reranked problems, accounting for exponential mapping from probability-based ratings to actual ranks.
2026-02-06: Removed 30k problems that had no pre-calculated solution. Fixed typos. Back button on linked problems goes back to challenge mode.
2025-09-27: Added a "Click to continue" screen to improve reliability when the app is backgrounded on mobile devices.
2025-09-26: Problem selection quality is now influenced by user solve rates.
2025-09-24: Adjusted ranks to better align with a human-calibrated AI model.
2025-09-23: The statistics page now tracks time spent in review mode.
2025-09-22: Changed how stats are tracked and presented so as to be more informative.
2025-09-21: Made problems easier by 3 ranks.
2025-09-20: Added another 25,000 problems. Made all problems harder by 1.5 ranks (an accidental adjustment in the wrong direction).
2025-09-19: Implemented automatic reload when a new version is available. Fixed a ghost move in review mode. Added 25,000 new problems. Adjusted ratings down by half a stone.
2025-09-18: Improved the score adjustment display and added new problems with ratings aligned to human network benchmarks. Adjusted the transition point for a better landscape mode experience. Fixed an issue where the about button was disabled. Removed unnecessary text from the sharing functionality. Added proper light mode support. Combined captures and komi into a single number. Added a legality check for ghost moves.
2025-09-17: Added progress bar for rank tracking. Implemented double-click confirmation on mobile devices.
2025-09-16: Added native sharing functionality with dynamic URLs. Fixed issues with mode switching between view, challenge, and review modes. Added visual warning for unreliable analysis results from API calls. Improved internationalization with consistent keys across all languages.
2025-09-15: The app install prompt is now shown only after solving a few problems (at 2, 19, and 190 problems solved).
2025-09-14: The main screen now always shows the rank of the current problem. Your personal rank is visible in the detailed stats view (by clicking the rank). Your rank change is now displayed between problems. Rank progression now slows down as you approach higher ranks.
2025-09-07: Problems are now randomly rotated for a fresh perspective.
2025-09-06: The total number of unique problems has increased to 41,000. You will now only be re-tested on failed problems that are easier than your current rank. Correctly answering a problem you have failed before no longer increases your rank.
2025-09-05: Added 10,000 new problems. Fixed a bug that could cause duplicate entries in your problem history.
2025-09-04: Improved the messages and animations when you fail a problem.
2025-09-03: Removed problems where the solution is to pass and problems with very unnatural moves. Improved handling of problems opened via shared links to not interfere with your history. Solving linked problems does not affect your rank.
2025-09-02: You can now see detailed statistics by clicking on your rank.
2025-09-01: Added 15,000 new problems. A warning is now shown if you try to load a problem with an incorrect ID.
2025-08-31: Added 5,200 new problems and removed the failure sound effect.
2025-08-23: Updated the user interface for a cleaner look and feel.
2025-08-22: The URL in your browser now updates to match the current problem.
2025-08-21: Your results are now anonymously stored on the server to help improve the application.
2025-08-20: Zooming on the page is now prevented to improve the mobile experience.
2025-08-19: Analysis mode is now available for reviewing problems.
2025-08-18: Illegal moves due to Ko are now marked with a square.
2025-08-17: Rank momentum increases more when you are on a winning streak.
2025-08-13: Sound effects for placing stones have been added.
2025-08-12: A fade-in/fade-out animation is now used for correct answers. Clicking on the komi value now shows the version number.
2025-08-09: The opponent's response is no longer shown on a correct move to reduce visual noise.
2025-07-29: The ranking algorithm for problems has been improved.
2025-07-28: The ruleset has been changed to Japanese rules with a 6.5 point komi. The application now uses zipped archives for problems, improving loading performance.
2025-07-27: The application is now a Progressive Web App (PWA) and can be installed on your device.
2025-07-26: The application now tracks the frequency of problems you have seen.
2025-07-21: The first version of the PWA went live.
2025-07-12: Initial commit.