Bidirectional translation

Bidirectional Translation

Activation Contract

Use for the active wave of language-loop, or for a standalone /learn drill [unit] session on a chosen unit (weakest-first when none is named). Requires an existing bilingual dialogue unit under .ai/learning/<topic-slug>/dialogues/.

Do not use as a grading instrument, for units the learner saw today (the delay IS the method — retranslation is spaced retrieval), or outside language topics.

Hard Rules

• Noticing over grading: the goal is to get curious about the differences between the learner's version and the original — not to score correctness. There is no pass/fail, no exam framing.
• Natural equivalence, never word-for-word: a different-but-natural rendering is a success; only differences that change meaning or break the target language's structure are material.
• From memory: the learner translates the unit's native text back into the target language WITHOUT looking at the original. Peeking mid-attempt just converts retrieval into re-reading — restart from the previous line instead.
• Delay is mandatory: drill only units captured in an earlier session (the language-loop N−5 offset, or any unit whose passive wave is at least a few days old).
• Classify every material difference into one category: word choice, structure/order, missing chunk, or grammar pattern.
• Every difference feeds exactly one loop (see Feedback Loops) — noticing without capture is wasted retrieval.
• Honest logs: record what the learner actually produced and noticed, in the unit's retranslation log.

Drill Protocol

1. Setup — open the unit's dialogue file; show the learner ONLY the native-language text.
2. Retranslation — the learner writes their target-language version from memory, line by line or whole, their choice.
3. Compare — place the learner's version beside the original target text; walk through the differences together, asking "what did the original do differently, and why might it?" before explaining.
4. Classify — tag each material difference with its category; ignore stylistic variation that is natural in the target language.
5. Capture — run the Feedback Loops below.
6. Log — append a row to the unit's retranslation log (date, differences by category, follow-ups created) and note the drill in path.md's pacing log.

Record the session with assets/bdt-session-template.md when run standalone (/learn drill); inside a language-loop session the unit's retranslation log row is enough.

Feedback Loops

• Missing chunk / word choice → anki-vocab batch candidate (register in vocabulary.md; Anki is its review system, no Leitner card).
• Grammar pattern / structure-order → spaced-recall card (cue: produce the pattern, not recite the rule).
• A pattern that also matches an english-tutor gap category → additionally append or reinforce the row in gaps.md so recurring gaps aggregate in one place.
• A unit with many material differences returns to the active wave later (re-drill after its next spacing interval); a near-clean retranslation retires the unit from drilling.

Output Contract

End every drill reporting: unit drilled, differences by category, follow-ups created (cards, vocab candidates, gaps rows), and whether the unit re-enters the drill queue or retires.

Attribution

Adapted from Luca Lampariello's bidirectional translation technique (L2→L1 at capture time, delayed L1→L2 retranslation, difference-noticing as the learning event), integrated with this repo's Leitner scheduling (spaced-recall) and Anki export (anki-vocab).