nx_convo.nx

nx_convo.nx -- WRITING arc, rung W-CONVO-1: the CONVERSATIONAL TUTOR core. Serves two operator asks with one engine: (1) "a couple doing language learning -- keep topics going" (2) "young people learn positive social skills to overcome anxiety, fun way" Three deterministic, data-driven primitives (the generation seat -- the LLM that actually speaks -- plugs in on top; this organ keeps the SESSION honest): cv_next_topic -- KEEP TOPICS GOING: from a topic bank (each topic has a difficulty + the turn it was last used), pick the eligible topic (difficulty <= level+window) that is LEAST RECENTLY used -> fresh, level-appropriate, never an immediate repeat. cv_rubric -- COACHING SCORE, RESPECT-GATED BY CONSTRUCTION: a weighted score across axes (warmth/authenticity/respect/confidence), BUT if the gate axis (respect / consent-reading) is below its minimum the attempt is DISQUALIFIED (-1) no matter how high "confidence" is. "Positive" is enforced in the math: you cannot win by being pushy. This is the anti-manipulation law. cv_level_adjust-- ADAPTIVE DIFFICULTY: nudge the target level up/down within bounds from the last score (overcome-anxiety pacing). All thresholds/weights/axes are operator DATA (rule 11), passed in -- nothing hardcoded. Pure integer, NO syscalls, NO imports, caller owns buffers. license_tier: ORIGINAL module: nishi-core.write.convo depends: capability: WRITE_CONVO_TUTOR