Methods

Elenchus is a retrieval-practice system (Karpicke & Roediger, 2008; Roediger & Butler, 2011; Dunlosky et al., 2013). A learner studies a card by trying to produce the answer, not by rereading the back. That choice is the whole product. Rereading and last-minute cramming are easy and feel like studying; they are weak for later recall (Dunlosky et al., 2013). Testing is not used here as assessment of a person. It is used as the study event itself — the same idea that has been shown in medical education as test-enhanced learning (Larsen et al., 2008).

Spacing is the second constraint. Reviews are delayed until recall would start to fail (Cepeda et al., 2006, 2008; Wozniak & Gorzelańczyk, 1994). Making retrieval slightly harder on purpose is the desirable difficulty (Bjork, 1994). The schedule is not a streak calendar and not a notification firehose.

Progress is reported as a retention score written in meters. The mountain photograph in the header is that number, not a badge for opening the app. Color starts at 1.00 m. A set’s full-mastery height is computed when the set is written, so ten decks show how much climb is still left if every card is kept.

The system runs in a browser on the learner’s own devices (a Windows PC, a Mac, a phone). Sign-in is optional for local study and used when the same library should follow the person. No classroom, no proctor, no sample recruited for a trial. The seeded exemplar is the January 2026 Texas Driver Handbook: official Appendix C items, chapter sets in one deck, plus sign recognition. Learners may add any other corpus.

A card has a front (the prompt), a back (the target), optional hint and media, and optional authored distractors used only in Learn (multiple choice). Drill is a flip. Write is fill-in-the-blank against the back. Distractors never pollute Drill or Write. When distractors are missing, the client prefers same-chapter answers of similar shape over shuffling unrelated backs.

Tokens

Knowledge volume is counted in tokens, not cards. A token is estimated as ceil(characters / 4) after whitespace collapse. That is the common portable approximation of a subword unit (see Sennrich et al., 2016), not a claim about any vendor tokenizer. A fifth-grader’s short fact and a board-exam paragraph are therefore not worth the same.

Density

Each set carries an ordinal density band: elementary 1.00, high school 1.35, college 1.80, professional 2.40, board exam 3.00. If the learner does not set a band, density is estimated from word length. These weights are a product convention so that harder material climbs farther per token. They are not a psychophysical scale and should not be cited as one.

Hold and meters

Held tokens = tokens × density × hold. Meters = held tokens / 24. One meter is therefore 24 everyday tokens still held, or eight board-exam tokens still held. Reported altitude is always shown to the hundredth of a meter.

The live header currently uses a stepwise hold: 0 if never reviewed, 0.18 if only seen, 1.00 if the card meets the retained rule (mastered, or interval ≥ 21 days with at least two successful reviews). That step is honest about “seen versus kept.” It is too generous about calendar time. The specified deterioration function, used to interpret a set a year later, is the FSRS retrievability curve described next — not a library-wide half-life.

Span

When a set is added, its span is the meters it would be worth if every card were retained today (hold = 1). The library span is the sum of set spans. The difference between span and current altitude is the climb still left. Ten sets do not “send someone into space”; they publish a ceiling. Whether the person is on that ceiling next year depends on review, not on the day the cards were pasted in.

Default scheduling is FSRS-5 (Ye et al., 2022). Memory is tracked as difficulty, stability S (days until predicted recall falls to 90%), and retrievability R. After each grade (again, hard, good, easy) the scheduler updates S and sets the next interval so that the review lands near the requested retention, default 90%. SuperMemo-2 remains available as a fallback (Wozniak & Gorzelańczyk, 1994).

FSRS-5 writes forgetting as a power function, not a single exponential:

R = (1 + (19/81) · t / S)^−0.5

That shape matches the finding that forgetting is rapid at first and then flattens (Wixted & Ebbesen, 1991, 1997). Ebbinghaus’s curve for unused nonsense syllables is much steeper (Ebbinghaus, 1885/1913; Murre & Dros, 2015) and is the wrong model for a handbook someone actually drilled. Under the FSRS-5 constants, predicted recall reaches 50% at about thirteen times current stability. A card that has only just become “retained” (S ≈ 21 days) is near half after roughly nine months without review. A card whose stability is a year is still near 90% at one year. There is no single “knowledge decays 50% per year” constant that the literature will support.

Long horizons need a floor. Bahrick (1984) found that Spanish learned in school dropped for about three to six years and then a remainder sat almost unchanged for decades — permastore — with the size of that remainder set by how well the material was originally learned, not by the calendar. The platform therefore treats a well-retained card as able to keep a residual hold (on the order of 0.15–0.25) rather than falling to zero. That floor is a reading of Bahrick, not a new experiment.

Well-designed computer tutors can approach the effect of a human tutor (VanLehn, 2011). Large language models add open-domain dialogue, which is useful and also risky: they can invent facts (Kasneci et al., 2023). Elenchus does not let the model be the source of truth. The source of truth is the card. The model is a tutor in the sense Mollick and Mollick (2023) describe as a coach or devil’s advocate — it asks, it does not replace the back of the card.

Connection is the Model Context Protocol (Anthropic, 2024). Claude, Grok, or Gemini on the learner’s machine can call a small set of tools against the local library:

• get_library — catalog, per-set span, missing choices, total climb left.
• get_context — the full set plus a brief for discussing the concept, not dumping an answer key.
• add_from_chat — turn a question the learner just asked into a card, with the question stored as provenance.
• set_choices — write three plausible distractors. Learn uses them. Drill and Write do not.
• due / review — the same FSRS grades the on-screen reviewer uses.

Prompts shipped with the server tell the model to generate distractors that a tired test-taker might actually pick, and to refuse to put those distractors on flip cards. When the learner is confused, the intended loop is: discuss the idea, then add a card from that confusion, then schedule it. The model is not asked to “know the handbook.” It is asked to work from the cards it just fetched.

Most forgetting happens between sessions. A tutor that only speaks inside the study tab is silent at the moment the schedule actually matters. The platform treats the language-model conversation as a second door into the same library. If the learner is talking to Claude or Grok about something else, the model can still read the due queue and the storm flag. That is the out-of-session channel that exists today: not a push notification, a tutor who is already in the room.

The decision rules follow just-in-time adaptive intervention design (Nahum-Shani et al., 2018) and the ordinary idea of a nudge (Thaler & Sunstein, 2008). Tailoring variables are already on the card: retrievability R, days since last review, count overdue, and the library “weather” (caught up, getting behind, overdue). Suggested rules, for a tutor that has just been opened:

• If weather is overdue and nothing has been reviewed in 48 hours: open with the smallest due set, one question, then stop.
• If a set’s held meters have fallen more than 10% of its span since the last session: say so in meters, not as a scold.
• If the learner is already talking about the topic: add a card from that question rather than changing the subject to “you have 40 due.”
• Never send more than one prompt per calendar day from the tutor. Never frame a lapse as a character failure.

Operating-system notifications are not required for this loop and are not claimed as part of the current method. If they are added later, the same rules apply: one timely ask, tied to a due card, easy to dismiss. A streak that punishes a missed day is not a nudge. It is a different product.

The meter is a bookkeeping identity, not a neural measurement. Density bands are assigned, not validated against exam scores. FSRS parameters are the published defaults until a learner has enough of their own reviews to fit. Language models can still hallucinate; grounding them in fetched cards reduces that risk, it does not remove it (Kasneci et al., 2023). Permastore is inferred from Bahrick’s decades-long Spanish data, which is recognition and recall of school learning, not a driver handbook and not a nursing exam. This document does not report outcomes. It reports how the platform is built, and which findings that construction is trying to respect.