Project status: Japanese immersion learning with Anki and selective glossing: 4. Deck hygiene: leeches and sentence-cued cards

Section 4 of 12 of the report "Project status: Japanese immersion learning with Anki and selective glossing" by Hiina (v1, updated 2026-10-27)
Report overview and contents: https://ai.hiina.space/reports/jp-learning-and-anki-tooling.md
Previous: https://ai.hiina.space/reports/jp-learning-and-anki-tooling/03-scheduler-settings-fsrs-6-on-anki-26.md
Next: https://ai.hiina.space/reports/jp-learning-and-anki-tooling/05-mining-from-more-anime.md

4. Deck hygiene: leeches and sentence-cued cards

Leeches ("rewrite the leech cards so they're less confusing")

Set the leech threshold to 6 with action Tag Only, as vnclub recommends. Cards then join a tag:leech rewrite queue instead of disappearing into suspension. Anki's default is 8 lapses with tag plus suspend (Anki manual: Leeches). The queue is worked with a four-way triage. The taxonomy is a synthesis, but each branch follows the manual or forum advice:

Interference with a similar word or kanji. Suspend one of the pair until the other is solid, or build a contrast card that shows both and focuses on the difference (forum thread).
Overloaded card. Several senses or a long definition. Shrink Lapis's MainDefinition to the sense actually met, and push the rest into Glossary.
Sentence-cued recall. Handled by the sentence-cue fix below.
Low-value word. Delete it.

When a rewrite changes the cue substantially, the card is Reset. Reset keeps review history, so the FSRS optimiser still sees reviews of the old cue (Anki manual: Studying). That is a mild contamination risk whose size nobody has verified.

For mnemonics on stubborn leeches, the companion memory-palace report supplies a constraint. Cues learners make themselves beat other-generated cues by a wide margin in lab studies (Mäntylä, via Wheeler & Gabbert 2017). A 30-student pilot also found that AI-generated associations far from the learner's own thinking can interfere with memory (J-STAGE). So in this project an LLM may propose a mnemonic or a contrast card, but the learner edits and chooses it.

Sentence-cued cards ("swap out the example sentences so cards cued by the sentence get reset")

This to-do has the best evidence behind it. A 2024 PNAS paper (six experiments on foreign vocabulary) found that varying the sentence cue across retrieval cycles beat a constant cue, immediately and after 24 hours, and more so when practice was spaced. Learners nonetheless judged constant cues more effective (PNAS via SWPS). That metacognitive illusion explains why sentence-cued cards feel fine until the word turns up somewhere else.

Lapis has four card types (word plus sentence, Click, sentence, audio) but no sentence rotation, and no add-on that rotates sentences per note turned up (Lapis README). That is absence in a limited search. The closest existing tool, anki-ai-dynamic-cards, generates a fresh LLM sentence at review time but supports English only and stores nothing (GitHub).

The manual sequence planned now:

Find the sentence-cued cards. Build a filtered deck of mature Lapis cards, temporarily flip them to a word-only front, and see which ones fail. Those are the sentence-cued ones (inference; this is a self-test, not a published method). This is the probe used in the next gate.
Switch every mined note to Click cards (IsClickCard), which hide the sentence until clicked, as Donkuri recommends (Donkuri mining).
Reset only the probe failures. Leave the passes alone so their stability survives.

Real rotation needs a custom template: an extra field holding several sentences, with JavaScript picking one at render time. Anki templates support JS, and a forum script already shuffles content on a card front (forum thread).

The extra sentences should come from real text, not free generation. The one Japanese-specific study found that all rater groups preferred retrieved corpus sentences over zero-shot generated ones (ACL SRW 2024). The project's own subtitle lines and Jiten's example sentences are the obvious corpus. Updating Lapis means importing a new APKG, which resets custom CSS (Lapis releases), so template edits are kept in a file that can be re-applied.

Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space