Bilingual Courseware — Terminology & Paired-Version Team

Maintains bilingual course materials with terminology discipline. Other skills produce the artifacts; this skill keeps their two language versions saying the same thing — same terms every time, same pedagogical demand on both sides, and a recorded pairing so that neither version quietly outruns the other. Default pair is EN ⇄ 简体中文 (the conventions file covers it deeply); any pair works with professor-supplied conventions.

Prime rule: the glossary is the contract. Translation without a confirmed glossary produces drift — the same concept rendered three ways across a semester, and 90 students learning the wrong standard term. Glossary-first is this skill's blueprint-first rule: no term-bearing material is translated until its terms are confirmed, and new terms encountered mid-translation pause for confirmation rather than being improvised inconsistently.

Quick Start

Build a terminology glossary for my machine learning course
Translate my syllabus into Chinese — the English version is the original
把这门课的教学大纲翻译成英文，专业术语要统一
I updated the English syllabus to v2 — bring the Chinese version up to date
检查我所有课件里的术语是否一致
Check whether my slides and exam use the glossary terms consistently

Modes

Mode dispatch rule: a translation request for an artifact whose domain terms are not yet covered by a confirmed glossary routes through glossary first — on the new terms only, incrementally, never a restart. Detect intent in any language.

Does NOT trigger

Agent Team (4)

Workflow (translate mode)

Phase 0  GLOSSARY GATE — check the confirmed glossary covers the artifact's domain
                         terms. Missing or stale coverage → `glossary` mode on the new
                         terms only (incremental — not a restart of the glossary)
Phase 1  DRAFT         — translator_agent produces the glossary-bound translation with
                         pedagogical-equivalence notes: examples that don't cross
                         cultures flagged with a proposed local equivalent; instructions
                         whose difficulty shifts in translation flagged
Phase 2  TERM AUDIT    — terminology_auditor_agent passes over the draft: every
                         confirmed term rendered exactly, near-misses are defects
         🧑 checkpoint: translation + divergence log — every deliberate non-literal
            choice listed with location and reason; the professor approves or redirects
Phase 3  REGISTER      — parallel_sync_agent records the pairing: which exact version
                         of the source this translation matches, in the sync register
                         and passport artifacts[]

glossary mode runs glossary_keeper end-to-end (extract → propose → 🧑 batched confirmation → version). parallel mode diffs the changed side, translates the delta through Phases 1–2, patches the pair, and updates the register; independently-evolved sides produce a drift report instead of a patch. check mode runs terminology_auditor_agent alone, read-only, findings located like submission-auditor.

Iron rules

1. Glossary-first. No translation of term-bearing material without confirmed glossary coverage. A new term encountered mid-translation pauses for confirmation — it is never improvised inconsistently, because the improvisation becomes the term students learn.
2. The professor confirms terminology. Discipline terminology is the professor's domain. Candidate pairs ship with an authority basis — standard textbook usage, national standards-body patterns, common EMI practice — but their status is [NEEDS PROFESSOR INPUT] until confirmed. A wrong standard term taught to 90 students is expensive; a confirmation batch costs minutes.
3. Pedagogical equivalence over literalism. Translated artifacts must make the same pedagogical demand, not the same words. Divergences are deliberate and logged — location, what changed, why — never accidental. Assessment translations are checked for difficulty shift: an exam that is harder in one language is an equity defect, not a style choice.
4. Version pairing is recorded. Every translated artifact records the exact source version it matches. Sync status is checkable at any time; drift is loud. The parallel mode exists because v2 of the English syllabus WILL forget the Chinese one — silence is the failure mode, the register is the cure.
5. Language-pair agnostic, EN⇄zh-CN tuned. references/emi_conventions.md covers the default pair deeply — register, punctuation, authority hierarchy, code-switching. Other pairs work with professor-supplied conventions, recorded per course; the skill never pretends its zh-CN depth transfers automatically.

Outputs

• bilingual/glossary.md — the confirmed terminology glossary (from templates/glossary_template.md), versioned, consumed by other skills
• bilingual/<artifact>_<lang>.md — translated artifacts, each with its divergence log
• bilingual/sync_register.md — pairing registry (from templates/sync_register_template.md): what matches what, at which version
• bilingual/drift_report.md (parallel mode, when sides evolved independently)
• bilingual/terminology_audit.md (check mode) — located findings, ordered by student impact
• Passport: paired artifacts cross-referenced in artifacts[]

References

• references/emi_conventions.md — EN⇄zh-CN register mapping, punctuation/typography, terminology authority hierarchy, assessment-translation equity, layout expansion, classroom code-switching conventions
• templates/glossary_template.md
• templates/sync_register_template.md
• Shared: shared/course_passport_schema.md (note course.language_of_instruction), shared/checkpoint_protocol.md, shared/quality_gate_protocol.md