Statistics

What a learner sees of their own study: a strip above the deck list that sums up the day, a statistics page for the whole instance, and a deck's own figures on its page. Everything is worked out in the browser from what the instance already holds in the pod; nothing is sent anywhere else.

The figures come from two places. The answer log is the history: every grade given, kept month by month. The review states say where each card stands now (srs.md). This page says what each screen shows and where its code is; how the figures are computed is on those two pages.

The today strip

Once the day has answers, a slim strip above the deck list (where a study session usually ends) says what the day came to:

• The streak leads, as what keeps a learner coming back: "3-day streak", then a note. On the first day it is "Day one! Come back tomorrow for two."; when the streak is the longest yet, "Your longest yet!"; else the record to beat, "Record: 12 days".
• A word of praise that grows with the day's answers: "Nice work!" under 10, "Great session!" from 10, "Impressive effort!" from 50.
• The day's numbers: how many answers, how many cards were new, and the share of today's reviews remembered with how many reviews it is of ("80% remembered of 5 reviews", below), then the week's answers so far ("12 this week"), with how they compare with the same days of last week when it had any ("+20% on last week"), then a link, "All statistics", to the statistics page. A day without a review (only new cards, repeats or course answers) leaves the share out. While cards are left to study today, due and new, the numbers say how many in every deck ("7 left to study today"), from the counts the deck list's rows read, once every deck's is known, and about how long they will take ("about 6 minutes to go", below), once how long an answer takes is known.

Before the day's first answer there is no strip. When the answers cannot be read there is none either: the deck list works without it. It reads the last 12 study months, whatever range the statistics page opens on, so the record is the year's; that is the page's own read when its range is the last 12 months, so the two never read them twice. Every change made clears that read, an answer in a session or a course among them, so the strip is up to date after a session; otherwise it is read again when the list opens and what was read is over half a minute old (below).

Code: TodaySummaryContainer.tsx reads the statistics and TodaySummary.tsx shows them; todayOf in statistics.ts picks the day out.

The statistics page

#/statistics?instance=…&range=…&view=… (routing.md), linked as "Statistics" from the instance bar on every screen of an instance, and from the today strip.

Above the figures, the range they are of: the last 30 days, the last 3 months (the default), the last 12 months or all time. Under it, four tabs: Overview, Activity, Memory and Cards. Both are in the URL (range, view), so a tab can be bookmarked or opened in a new browser tab, and a change of either replaces the history entry: Back leaves the page rather than walking its tabs. The tabs are links, and a tablist: the arrow keys, Home and End move along them; on a phone their strip scrolls sideways in its own box. Before any answer the first three tabs say that the statistics begin with the next study session; the Cards tab shows where the cards stand anyway. A range with no answers, in a log that has some, says so instead, and that a longer range may show earlier ones.

The figures that run up to today look back as far as the activity calendar, its 52 weeks (or the whole log, for all time), whatever the range, so a short range does not cut them short: the streaks, the days studied of the last 30, the calendar, what each day held, this week and month beside the last, and the new cards beside the reviews by week. The rest are the range's.

A note comes first on the overview, activity and memory tabs, "Since March 2026, your first month of study.", when the answer log begins inside the range read (always so for all time): the figures that need the answers before them, such as how long cards took to grow mature, cannot reach further back. Each tab, from the top:

A card is young while it is reviewed at an interval under 21 days and mature from 21 days on. A card forgotten last time and not remembered since is relearning, whatever its interval: SM-2 leaves it no repetition in a row (srs.md), as the Studio's "learning" state does; it is no beginner's stage, as a card passed once is not back in it until it is forgotten. SM-2 spaces a card's reviews 1 day, 6 days, then about 15 and about 38, growing from there, so some bins of the interval chart stay nearly empty; its caption says so. A card studied both ways counts once each way, and the page says so. A share is always shown with the count it is of, however small, so that "100%" of one review reads as what it is.

Every chart can be read without seeing it. Pointing at a day, a weekday, a week, a month or a part of a bar (or tapping it) puts what it held in words under the chart, and "Show as a table" lists the same figures. The keyboard reaches the same words: each chart is one tab stop (each bar of the answer buttons with answers is one), Left and Right step along its points, Home and End go to the first and the last, and Escape lets go; in the calendar, Left and Right step a week and Up and Down a day (usePoint). The line under the chart is the chart's description, so a screen reader that reaches the chart says how to step through it; keys pressed with Alt, Control or Meta are left to the browser. A point reached by the keyboard is said once more in a status line, so a screen reader reads each step out, while sweeping the pointer stays quiet. A screen reader that reads the page in browse mode keeps the arrow keys for itself: its focus mode passes them on, and the tables say the same. Only shades of one green are used; where two lines meet, a dash or a legend tells them apart, and where bars hold several series, a legend names each step of the green. In forced-colors mode the steps keep apart as empty, hatched and filled. The histograms are the Studio's own, shared from the UI package.

The page reads three things on their own, so none waits for another: the answers of the range (UseCases.getStatistics); the cards and review states of every deck with, for what is hard, the last 12 study months of answers, whatever the range (UseCases.getCardProgress); and, with a course among the decks, how the courses have gone (UseCases.getCourseStatistics), from the same 12 months and each course's outline, which the library reads once a page load. A course whose outline cannot be read is left out. All three ask for the months at once, and a month already being read is read once for all (unless an answer was written meanwhile, which that read may miss); after that, what is kept serves each. The months are read beside the decks' cards, and only what is hard needs them: when they cannot be read, what is hard is left out and where the cards stand still shows. When the cards cannot be read, a quiet line says so and the rest of the page works on; when the courses cannot be, they are left out.

Loading

The answer log is the large part: one to three megabytes of Turtle a month, by server, for a learner who answers a hundred cards a day. So:

• Each month is read once a page load, then only when it changed. getStatistics and the Studio's loadAnswerLog share one cache of the months read, by their ETag: a month already read costs a 304. A read of some months adds to it, never replaces it, so the 12 months read for the statistics and the whole log read by the Studio do not undo each other; a month the log no longer lists is dropped. The month being studied changes with every answer, so it is read again after studying. A server that sends no ETag (Node Solid Server) is read whole every time.
• What was read is shown again for half a minute (staleTime) on the statistics page, a deck's page and the today strip, so moving between them reads nothing. Every change made clears the statistics, the courses' figures among them, and where the cards stand at once, as it ends (whether it was written or not): an answer, a card or deck added, edited or removed, a deck's pace, the preferences, a guest's study brought over, a Studio edit. Each is a mutation, and the app clears them after every mutation (keepStudyFiguresFresh in studyFigures.ts), so a new kind of change cannot be left out. A change made on another device shows once what was read is half a minute old. The courses' figures are also read again when a chapter is completed, as the chapters completed are part of what they are kept by.
• The range read. getStatistics reads the last 12 study months by default, or range: the last so many study days, or the whole log ("all"). It also reads the activity calendar's 52 weeks (calendarStartOf) when the range begins after them, for the figures up to today: statisticsOf counts the range's answers from its first day (since), and those figures, with recentDays, from every answer read. The months read are the same the today strip's read takes in, so they come from the shared cache. Its result's coverage is the log's first month when the log begins inside the range, and logBegins the log's first month, for the empty range's message. The page's ranges are 30 days, 90 days (its default), getStatistics' own 12 months and "all" (rangeRead and statisticsQuery in studyFigures.ts), each read once and kept apart; the last 12 months are kept under the today strip's key, so the strip and the page share that read. All time downloads the whole log once a page load, as the Studio's loadAnswerLog does, and on a server without ETags at every visit.

Measured against a real server (history.integration.test.ts, run with MEASURE_ANSWER_LOG=1; testing.md), for 100 answers a day, on a development machine with the server on it:

Parsing costs about 27 µs an answer whatever the server writes; Community Solid Server writes every IRI in full, so it sends almost three times the bytes. Read again, the log costs a request a month and no parsing.

A deck's own figures

A deck's page (#/deck) sums up its figures of the last 3 months, once it has answers in them: the young and mature recall tiles, recall each way when it has reviews both ways, first try and, with course answers, the graded course answers tile (alone for a deck studied only in a course's chapters). Under them, always, "More statistics", a link to the rest, whose longer ranges reach a deck's older answers (DeckStatisticsContainer.tsx). That read is the deck's statistics page's at its default range, so following the link reads nothing again. Then where its cards stand, all new before the first answer, with a tile for the day its new cards run out (none for a course's deck, or once none is new). Then what is hard in it, from its answers of the last 12 study months (HardestCardsContainer.tsx): "Hardest cards", the ten cards in use forgotten most often in their reviews, ranked as the leeches are, a leech marked as one, each a link to its page (none for a deck without reviews); and, for a deck with course answers, "Most-confused options": the wrong options chosen most often, each with its card and how often (confusionsOf).

The deck's statistics page, #/deck-statistics?instance=…&deck=… (under the deck in the breadcrumbs: Decks › deck › Statistics), has the range and the tabs of the statistics page, of the deck's answers alone: its own activity calendar on the overview, without the table by deck; recall each way on the memory tab when it has reviews both ways, with its hardest cards and most-confused options in place of the leeches across decks; its own cards and, for a course's deck, its course on the cards' tab. Its download holds the deck's answers only.

A card's history

A card's own page (#/card) ends with its history (CardHistory.tsx): for each direction it has a review state, its interval, its next review and the study day it was first studied on (by the day's boundary hour, as the answers' days are); then every answer to it in its deck, from the whole answer log (UseCases.cardAnswers, as the Studio's card inspector reads it), newest first, as a table: the study day, which way (when the card is asked both ways), what kind of answer it was (a first answer, a review, again the same day, or a course question), the button pressed (Again, Hard, Good or Easy, as the answer buttons count them) or, for a course question, right or wrong and the wrong option chosen, and the interval it gave. A card not studied yet says so.

What the figures count

Recall, lapses and course answers need more than one answer to tell: whether an answer repeats another to the same prompt that day depends on the prompt's earlier answers. answerChain.ts groups the answers by prompt (deck, card and direction) in the order they were given, once for each set of figures, and the definitions below read those chains.

Growth and depth of memory

memoryCurve.ts reads the same chains for how memory grows. The log holds the answers only, not the Studio's edits, so what it replays is approximate, and the page says so.

The range's figures are of its answers, but lateness, where the cards stood and the time to mature also read the answers before the range that the page reads anyway (the activity calendar's 52 weeks, or the whole log), so a review early in the range still has the answer before it. Course answers change a card's schedule, so they count in where the cards stood, but never in recall.

Time spent and the hours of the day

An answer before vocabulary 1.19 records when it was given, not how long it took. Since then an answer records its own response times and the device's offset from UTC (data-model.md, below). studyTime.ts counts the times recorded from the first answer read that has one, and estimates the time before it from the pauses between answers. Every figure of it says "about", as recorded times are rounded and the rest estimated.

Response times

Since vocabulary 1.19 each answer says how long it took and where it was given, measured by the screen that asked it and kept in the answer log (data-model.md). Nothing is asked of the learner and nothing turns it off; privacy.md says so.

Why the figures changed

Before this, recall and "remembered" counted every answer with a prior interval, and a lapse was any answer below grade 3. That made them read better and worse than the learner's memory was:

• A card failed and put back in the session is answered again minutes later, with a prior interval of a day, and usually right. Each such repeat counted as a young review remembered, so young recall, and today's "remembered" most of all, came out too high. The strip's share was of every answer of the day, first answers and course answers included.
• Course answers are chosen among options, graded 3 or 1, and only partly logged; they counted as recall.
• A card failed when it was introduced, and every repeat failed the same day, counted as a lapse, so leeches came early, in the Studio too.

The figures a learner sees therefore changed with this definition: recall and the strip's share are of scored reviews only and usually lower for young cards, course answers have their own figure, and lapses and leeches count review lapses only, so the Studio shows fewer. Nothing in the pod changed: the same answer log is read differently.

Where the code is

The statistics are the same for a guest (guest-mode.md), whose answers are kept in the pod in their browser and move with the rest of their study when they log in.

How it is tested

• Unit tests at 100% coverage, as everywhere (testing.md): answerChain.test.ts, statistics.test.ts, memoryCurve.test.ts (a forget, a move with its progress, a reschedule, introductions before the answers read), studyTime.test.ts, answerCsv.test.ts, studyPatterns.test.ts, cardProgress.test.ts, cardHistory.test.ts, scheduleInsight.test.ts and course.test.ts in the domain, the passes over the whole log (chainIndexOf, statisticsOf and the answer buttons, first try, recall each way, where the cards stood, the estimated time and the hardest cards) held to linear growth, the use cases in useCases.test.ts, and StatisticsScreen.test.tsx, StatisticsSections.test.tsx, router.test.ts, Workspace.test.tsx (the range and tab in the URL, the deck's statistics route), the chart tests in charts/, CardProgressContainer.test.tsx, HardestCardsContainer.test.tsx, CourseStatistics.test.tsx, CardHistoryContainer.test.tsx and TodaySummary.test.tsx in the UI; StudyContainer.test.tsx and ChapterPlayerContainer.test.tsx check that an answer clears the statistics.
• Against real servers: history.integration.test.ts checks that the statistics and the whole log share the months read, a month read again only once it changed, that lateness, where the cards stood and the time to mature follow a card across months (journeys cannot span days), and measures a large log when asked (above).
• User journeys: in full-learner.journey.ts, step 13 checks that the today strip counts each session's answers as soon as it ends, and step 15 opens the page after two decks were studied, one card answered Again and then once more, and checks the tiles (the repeat is no review), the note on where the log begins, the answer buttons (the Again among the first answers, the repeat left out), recall by interval (every span with no review yet), this week's young and mature cards, what the day held (one repeat), the weekdays, the time spent (its sessions a count, never its minutes, which hang on how fast the journey runs) and the answers by hour of the day (their total), the table by deck, every maturity bar (relearning among them), the overdue tile, the interval chart and that no leech is listed before any review, then reads today in the calendar from the keyboard, and over all time that the time studied is given, recorded since today; step 15b reads the history of that deck's two cards, the repeat told apart. In course-chapter.journey.ts, step 09 checks that a course's answers show as graded course answers and not as recall, and the chapter just completed under "Your courses", done, with its graded answers; and step 10 that the course's deck names the wrong options chosen in its final review. In studio.journey.ts, steps 30 and 32 check that trying a course in the Studio's sandbox leaves the statistics as they were. The page object is Statistics.ts. The latest run's screenshots are online at https://solid-memo.com/journeys/.