How we verify

1. Draft extraction (draftFromSource)

The input is the plain text of an imported announcement. Fixed rules pull out the fields: the first line becomes the title; keyword matching finds the audience, a date or deadline, and a contact; list items under the “required documents” and “steps” headings become those lists; every field is length-capped.

It does not rewrite sentences and it does not add anything the source does not contain. If a field is not found, it stays empty.

Risk is always set to medium at this stage; a human assigns the real level against the definitions above.

The return value always carries the “requires human review” flag. No input can make it false.

2. Question answering (answerFromApprovedContent)

It retrieves from already-approved content and generates no new claims.

A question passes three keyword rules first. Anything touching document numbers, passwords or verification codes, and anything we have no basis for — a scholarship outcome, a guarantee, what will happen next year — returns “not enough information”. Anything legal, medical, visa-eligibility or emergency related returns “high risk, use an official channel”. Only what is left returns approved content.

In other words, it is built to say it does not know rather than to guess.

Access gates

• The draft endpoint (POST /api/ai/draft) requires a signed-in administrator and an explicit consent flag on the request; anything else is rejected with a 400.
• The answer endpoint (POST /api/ai/answer) requires a signed-in account.
• The source import endpoint (POST /api/source/web) requires an administrator.
• Public pages read the approved content files, which are written to version control only after content review. There is no automatic path from a draft to a public page.

Limits of this section today

• Human review currently happens by editing the approved content files in version control. The side-by-side source-versus-draft review screen does not exist yet.
• Pages do not yet carry a per-page content-state badge saying whether a page was AI-assisted or written by hand.

Interface languages

中文 · English · Bahasa Indonesia · Tiếng Việt

Content languages

中文 · English

Interface strings are the navigation, the buttons, the field captions, the section headings and the empty states. The Indonesian and Vietnamese interface strings have not yet been read by a native speaker; the review list is in the repository's content-review directory.

The 16 tasks have cleared human content review in Traditional Chinese and English only. An Indonesian or Vietnamese page shows a machine translation of that English edition which no native speaker has read; every such page says so before it shows the content, and prints no verification date.

Why these four

• The Ministry of Education publishes open data on international-student numbers per institution, as a row-level file. This site's author aggregated that file to get a nationality breakdown for this university. NTUST publishes no such breakdown, so this is our own aggregation, not a figure the university released.
• The method: filter the ministry's row-level file to this institution's degree-seeking rows, total them by nationality, and sort. This page quotes no headcount — that aggregation has not been re-derived here, and we do not repeat a number we have not re-checked.
• What decides a language is not the headcount by itself but whether an existing edition is readable. Among the largest groups, Indonesian and Vietnamese students are neither Chinese-reading nor studying through English, so each added interface language adds readers.
• The languages left out were left out for reasons too: groups that are mostly 僑生 already read Chinese; groups that study through English already have the English edition; and where a country has single-digit degree enrolment, an interface language adds no reader.

Dataset

Ministry of Education | international students at higher-education institutions (data.gov.tw) （opens in a new window）

What we do not do

We do not publish UNLABELLED machine translation of guidance on residency, health insurance, tax or medical care. When one of those pages is translated wrongly the cost is a missed legal deadline, and the reader has no way to detect it. The Indonesian and Vietnamese editions ARE machine translations today, so each of those pages states it before the content: produced by AI, not yet read by a person, verification record held by the Chinese and English editions — with links to both reviewed editions and to the cited sources. Until a native speaker has reviewed them, those pages stay noindex.

If something is wrong, out of date, or different from what you were told at the counter, report it. This is our main way of finding content that has drifted.

A report is stored with an open status. When an administrator resolves it, an audit record is written saying who resolved which report and when.

Limits today

• Reporting requires a signed-in account and a note of at least 10 characters.
• The audit trail is currently visible only in the admin area; the public changelog page is still being built.
• We do not promise a response time. Better to say that than to publish a number we cannot keep.

The content is compiled from public official sources; keep the source links when you quote it. Our own licence terms and contact address are not decided yet — the fields below are unfilled.

Content licence

[TO FILL: content licence]

Corrections and contact email

[TO FILL: contact email]

Back to About us