Anki's algorithm scales fine. The rating signal does not.

No. The algorithm is well-behaved at that scale. FSRS fits a per-card DSR triple (Difficulty, Stability, Retrievability) and re-fits parameters when you run the optimizer; the cost per scheduling decision is O(1) per card and O(N) over the deck during optimization. A 30,000-card collection on a modern laptop optimizes in seconds. The user-facing problem at scale is review volume, not algorithmic cost. Daily reviews settle at roughly 10x your new-cards-per-day rate (Issa Rice's c·log_a((a-1)D+a) formula), so a med student adding 80 to 120 new cards a day during pre-clinical years lands at 800 to 1,200 reviews a day by year 2.

It comes from integrating the review schedule under SM-2-like spacing. If you add c cards per day and each card is reviewed at intervals of a, a^2, a^3 days (with ease factor a around 2.5), the total cards due on day D works out to c·log_a((a-1)D+a). The function grows logarithmically without bound: 5 cards/day at a=2.5 produces 5 reviews on day 1, 42 by year 5, 50 by year 20, 56 by year 50. The practical 10x rule falls out of that math once D is large. FSRS is more efficient than SM-2 at the same retention target, so the multiplier is usually a bit lower than 10x in practice, but the logarithmic shape is the same.

Available study time per day does not grow with deck size. If you have 90 minutes to do Anki in the morning and your due queue is 200, you have 27 seconds per card. If the queue is 800, you have 6.75 seconds. At 6 seconds you can scan a stem, recognize wording you have seen before, and rate Good. You cannot read a clinical vignette, run a differential, eliminate two distractors, and decide between the remaining two. The algorithm does not know which kind of retrieval just happened; it only sees the rating button you pressed.

Because at low time-per-card budgets, the only retrieval the brain has time to run is verbal pattern-matching. Pattern-matching succeeds when the stem is static, so you rate Good in 1 to 2 seconds. FSRS updates Stability upward, extends the interval, and the deck statistics report higher and higher retention. The exam stem reads differently from the deck stem (different patient age, different presentation, different distractors), the pattern-match key does not fire, and you fail. The algorithm did exactly what you told it to do. You just told it the wrong thing for several thousand reviews in a row.

No. Raising desired retention shortens intervals and gives you more reviews per day on the same deck, roughly 1.5x to 3x more depending on the deck. If your rating signal is dishonest (you are pattern-matching static stems), more reviews of the same dishonest cards just compound the false-mastery feedback faster. The Anki docs flag this in deck-options.html: the retention slider trades off review burden against forgetting probability, it does not change what the rating means. The fix has to happen upstream of FSRS, where the card surface lives.

Two parts. (1) Card-quality controls before emission: source-anchored stems (every card traces to a slide or page), length-matched distractors (within ~25 percent character length), grammar parallelism (no a/an tells), no filler templates (no 'all of the above', no 'both A and C'). (2) Stem rephrasing on each surfacing: the same underlying concept appears in different verbal surfaces (original wording, paraphrase, clinical vignette, analogy) so the pattern-match cache never fires twice on the same key. Studyly handles both as part of the generation and review pipeline; static .apkg files imported into Anki get only the first part, which is why a periodic re-export is the workaround when reviewing in Anki rather than in Studyly.

FSRS Optimize re-fits the 17 weights of the DSR model against your review log via gradient descent. It runs in seconds even on a collection with hundreds of thousands of reviews. Anki recommends running it after roughly 1,000 reviews of new history and after any large import. At med-school scale a monthly cadence is plenty; running it after every study session is wasted effort. The optimizer cannot fix the rating-signal problem either; it just fits the DSR triple better to whatever signal you have given it.

You can, but the math is unforgiving. The 10x rule (reviews per day at steady state ≈ 10x new-cards-per-day) means 200 reviews/day requires you to cap at 20 new cards/day. A pre-clinical block with 30 lectures a week and 50 to 100 testable facts per lecture produces around 1,500 to 3,000 facts a week. Twenty new cards a day is 140 a week. You will be permanently behind. Capping new cards solves the volume problem by trading away the coverage problem; that is a different failure mode, not a solution.

When the same concept surfaces with rotating verbal surfaces, your retrieval has to traverse the concept on each surfacing instead of hitting a cached stem-to-answer mapping. The wall-clock time per card goes up (8 to 12 seconds instead of 2 to 3) but the rating you give back actually reflects concept retention. FSRS then sees a more honest stream of ratings: Good means 'I retrieved the concept', not 'I recognized the wording'. The interval growth tracks real retention. On the held-out three-document eval, Studyly's cards score 81.3 on factual correctness, clarity, distractor quality, and question-type coverage (vs Unattle 78.0, Gauntlet 68.0, Turbolearn 57.8). The eval is orthogonal to FSRS but it measures the upstream layer.

Not in any practical sense. People run 100,000-card collections in Anki without algorithmic issues; the FSRS model is happy with that. The user runs out before the algorithm does. The two soft ceilings in practice are (1) time-per-card collapsing below the retrieval-honesty threshold (around 8 to 10 seconds per card for non-trivial MCQs) and (2) deck heterogeneity making the DSR fit noisier (mixing 5,000 anatomy cards with 5,000 pharmacology cards in one preset gives a less personal fit than two presets). Both are organizational problems, not algorithmic ones.