bidirectional-translationlisted

# 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: