1

For an @-subcard during a local quiz, the prompt sentence will be shown but with a piece of it replaced with the @-subprompt. In order to know which piece of the prompt sentence to hide, the @-subcard's pieces will be examined for cloze marks.

if no cloze marks found, look for the prompt, and if that fails, look for the responses, and if that fails, throw.
if cloze marks found in one item (or more, though that's probably an error), hide that in the sentence.

We can have a "linter" that will warn when no clozes are present in @-subcard and the prompt/response chosen is repeated in the sentence. Also if more than one item has a cloze mark.

2

Two Ebisu models are associated each global card. One global one shared by all instances of it. And one local one personal to it and this sentence. When you do a global review (all by itself), the local ones all get an active update. When you do a local review (with a single sentence), the global get a passive update.

3

Exponential decay and SRS might optimize long-term just-about-to-forget, hard-to-recall recall, but what if you want to optimize rapid recall? You see kanji on the subway, and sure if you've been doing your SRS you'll likely eventually get the answer (with a probability given by Ebisu), but you'll have rolled past it. The only way to do that is regular exposure.

fasiha / curtiz-japanese-nlp

Notes #1

1

2

3