Skip to main content
Every course the Producer agent curates deposits its glossary into one shared store. It isn’t a per-course word list — it’s a cross-document base that gets thicker and more cross-referenced every time a new course is produced.

​
The concept-variant model

A term isn’t stored as a single zh↔en pair. It’s stored as a concept — a canonical Chinese form, normalized so that simplified/traditional variants collapse to the same row — with one or more term variants attached to it, each tagged by language and a status: preferred, admitted, or deprecated. A concept can have several admitted English renderings while one is marked preferred; a deprecated variant stays in the database (so old occurrences still resolve) without being surfaced as current usage. Every place a term actually appears in a course text is a separate occurrence row, linked back to its concept variant with the exact character span in both the Chinese and English paragraph. That’s what makes hover-linkage and highlighting possible: the reader isn’t guessing where a term is on the page, it’s rendering spans that were recorded at production time.

​
Official-tier grading

Not every rendering in the base carries the same authority, and InterpScout doesn’t pretend otherwise. Two places do the grading work:
  • During course production, the adversarial glossary reviewer sub-agent’s job includes flagging renderings that are plausible paraphrases but not the established, official form — the Producer agent is expected to verify uncertain renderings with a web search rather than write from memory.
  • In briefings, every entry in the bilingual term table carries an explicit two-column verdict: a grading column (official vs. reference) and a source column. A rendering is only marked official if it was actually found in an authoritative bilingual source — a government white paper, a Ministry of Foreign Affairs bilingual transcript, a UN parallel text. Anything the agent can’t source that way is labeled reference, honestly, rather than upgraded on confidence alone.

​
Search: literal, phonetic, and semantic, fused

Searching the terminology base runs three matching strategies against the same query and merges the results:
  • Literal substring matching on Chinese and English text — fast, exact, and the primary path for most queries.
  • Pinyin matching, so a query typed in pinyin still reaches the Chinese term it’s meant to find.
  • Semantic vector search over term embeddings, for queries that don’t share surface text with the term they’re looking for.
Results from these paths are fused with reciprocal rank fusion and then reranked, so a query gets one ranked list rather than three disconnected ones. Search results are term hits, not course listings — searching 新质生产力 returns the term itself, its rendering, and its grading, ranked by relevance, not a list of courses that happen to mention it.

​
Reading a term in context

In the parallel reader, a highlighted term is a live link into this model, not decoration:
1

Hover

Hovering a highlighted term on one side of the parallel text highlights its counterpart on the other side — the zh↔en linkage is real, driven by the recorded span, not a heuristic guess at alignment.
2

Double-click

Double-clicking a term jumps to that course’s own glossary table, anchored at the row for that term, showing its graded rendering and source. There is no popover card floating over the text — the glossary lives at the bottom of the page as a real, readable table, and double-click scrolls you to the right row in it.
Grading and the concept-variant model are course-terminology-base concepts. Briefing term tables are meeting-specific by design — their terms are graded the same way but stay scoped to that briefing rather than merging into the long-lived cross-document base, since a one-off meeting term isn’t necessarily a durable asset for the glossary.