Study in China Atlas

Source-led discovery for international students exploring Chinese universities, programs, scholarships and study cities.

Explore the live atlas · Browse programs · Public API · Data policy

Important

Study in China Atlas is an independent, non-commercial public-interest directory—not a university, scholarship provider or application portal. Always confirm deadlines, fees and eligibility on the linked official source before applying or paying.

Why this atlas exists

International applicants often have to compare hundreds of differently structured university pages, PDF notices and application systems. Study in China Atlas turns those scattered official sources into a searchable, multilingual catalogue while keeping the original evidence one click away.

The product is built around three promises:

Discover broadly	Verify precisely	Understand uncertainty
Explore a national selection instead of a handful of famous universities.	Dynamic facts carry an official source, check date and publication status.	Unknown, conflicting or stale values stay empty; old cycles and guesses never fill the gap.

Public catalogue snapshot

266 universities	1,234 programs	356 scholarships	62 cities

The public catalogue also contains 256 published admission-cycle records and is backed by 2,070 registered official source records. Snapshot evaluated for 2026-08-10 with npm run quality:platform-scorecard.

Open the honest data-depth scorecard

Record count is not the same as record completeness. These are the current depth and platform-readiness indicators:

Quality indicator	Current baseline	Next gate
Universities below three published programs	8	0
Programs with a current public cycle	253 / 1,234 · 20.50%	≥ 70%
Programs with duration	61.59%	≥ 90%
Programs with an official application route	50.89%	≥ 80%
Programs with known teaching language	85.09%	≥ 95%
Programs with eligibility/language evidence	6.00%	≥ 50%
Universities connected to scholarships	207 / 266	≥ 230
Cities with reviewed coordinates	27 / 62	62 / 62
Source Manifests registered	10 / 266	266 / 266
Completed V2 Source Manifests	0 / 266	266 / 266
Complete catalogue reconciliation	0 / 266	266 / 266
Platform quality gates passing	3 / 14	14 / 14

The raw compatibility dataset contains 272 universities, 1,255 programs and 384 scholarships. Draft, archived, identity-conflicting or publication-ineligible records are intentionally excluded from the public numbers above.

What applicants can do

• Browse universities, programs, scholarships and student cities in one coherent interface.
• Explore programs through a 17-field, applicant-oriented taxonomy, including Chinese language and international Chinese education.
• Share URL-based filters, sorting and pagination; remove active filters individually and preserve browser history.
• Move from application state and deadline to fees, language and duration through decision-first cards and detail-page application snapshots.
• Narrow programs to records with an explicit university or program scholarship relationship without claiming applicant-specific eligibility.
• Explore cities through a geographic constellation or an accessible searchable directory, then use flagship guides with official sources, FAQs and stable section links.
• Keep identity-only records available for official discovery while excluding thin pages from search-engine indexing until they meet deterministic completeness gates.
• Save records locally, compare up to four programs and print a compact comparison sheet.
• Inspect field-level source links, last-check dates and uncertainty states.
• Use English, Chinese and Russian public routes; German, French and Spanish are the first reviewed expansion routes.
• Move from discovery to the official university or scholarship application system—the atlas never receives application documents.

Portuguese and Arabic remain registered preview locales and are not publicly indexed. When translated record prose is unavailable, the interface uses an explicit English fallback rather than presenting machine output as an official translation.

Trust model: evidence before completeness

The catalogue separates record identity from field visibility. A verified program identity may remain discoverable even when its tuition or deadline cannot safely be shown. Each dynamic field uses one of six explicit states:

type FactStatus =
  | 'known'
  | 'officially_not_announced'
  | 'not_applicable'
  | 'source_unavailable'
  | 'conflict'
  | 'stale'

Publication follows these rules:

1. Only allowlisted university, government and scholarship-provider sources can support a fact.
2. Every changing fact belongs to a specific academic year and intake.
3. Deterministic parsing extracts links, dates, money and page structure before model-assisted extraction.
4. High-risk fields require two independent MiniMax extractions to agree and point to locatable evidence.
5. Evidence, deterministic rules, freshness and cross-source conflict checks must all pass.
6. A failed field is published as an empty value with status metadata and an official entry link—not as a guess.
7. Immutable releases are relationship-, count- and hash-validated before the public pointer moves atomically.

Raw HTML, PDFs and screenshots stay private in R2. The public platform exposes only structured facts and the shortest necessary source context.

System architecture

flowchart LR
    S["Official university, government and scholarship sources"]
    Q["Cloudflare Queues"]
    I["Ingestion Worker<br/>fetch, parse, dual validation"]
    RS[("Private R2 snapshots")]
    P[("Pipeline D1<br/>claims, evidence, quarantine")]
    E["Entity Materializer"]
    U["Publisher"]
    B["Release Builder"]
    RR[("Versioned R2 releases")]
    C[("Catalog D1<br/>immutable public projection")]
    A["Catalog API"]
    J["Generated JSON compatibility snapshot"]
    R["json / shadow / d1 Repository"]
    W["Next.js 16 on Vercel"]

    S --> Q --> I
    I --> RS
    I --> P
    P --> E --> P
    P --> U --> P
    P --> B
    B --> RR
    B --> C
    C --> A
    A --> R
    J --> R
    R --> W

Loading

Why the platform is split this way

Layer	Responsibility	Engineering reason
Pipeline D1	Jobs, candidates, evidence, conflicts and quarantine	Frequent internal writes never compete with public reads.
Private R2	Compressed source snapshots, PDFs, evidence assets and release exports	Cheap immutable storage makes audit and recovery possible.
Catalog D1	Validated, versioned public projection	The website reads a stable release instead of half-written ingestion state.
Catalog Repository	json, shadow and d1 backends	The application can compare backends and roll back without rewriting pages.
Next.js on Vercel	Multilingual pages, SEO, feedback and web delivery	App Router provides server rendering while Vercel supplies previews and production delivery.

The platform foundation is implemented, but production remains deliberately compatible with the generated JSON repository while D1 shadow comparison and release-readiness gates continue. MiniMax credentials exist only as Cloudflare secrets; source content never receives database, network or execution privileges.

Technology map

Concern	Choice
Web application	Next.js 16 App Router, React 19, strict TypeScript 5.9
Design system	Accessible custom “academic atlas” CSS, progressive enhancement
Collection	Cloudflare Workers, Queues, allowlisted fetch policies
Data and evidence	Pipeline D1, Catalog D1, private R2, immutable releases
Extraction	Deterministic parsers + MiniMax dual extraction + evidence grounding
Validation	Zod schemas, deterministic conflict/freshness/publication gates
Testing	Vitest 4, Testing Library, Node test runner, Playwright
Delivery	GitHub Actions, Vercel Preview and Production deployments
Observability	Release age, queue/DLQ state, source health, backup and cost signals

Public API

Versioned endpoints use cursor pagination. List endpoints default to limit=24 and accept at most 100 records per request.

Endpoint	Purpose
GET /api/v1/institutions	Search and filter institutions
GET /api/v1/institutions/{slug}	Institution detail
GET /api/v1/programs	Search and filter programs
GET /api/v1/programs/{slug}	Program detail
GET /api/v1/programs/{slug}/cycles	Program admission cycles
GET /api/v1/scholarships	Search and filter scholarships
GET /api/v1/scholarships/{slug}/cycles	Scholarship cycles
GET /api/v1/releases/current	Current public release metadata
GET /api/v1/double-first-class	Double First-Class coverage view

curl "https://studyinchina.vercel.app/api/v1/institutions?discipline=chinese-language&limit=3"

Program filtering supports institution, city, program type, degree, field, teaching language, academic year, intake, tuition range, application state and scholarship availability.

Run locally

The default JSON backend works without Cloudflare credentials, so a contributor can inspect the product safely before configuring infrastructure.

git clone https://github.com/computersciencefreshmen/StudyInChina.git
cd StudyInChina
npm ci
cp .env.example .env.local
npm run dev

PowerShell equivalent:

Copy-Item .env.example .env.local
npm run dev

Open http://localhost:3000. The root route redirects to the saved or browser-preferred public locale.

Set CONTENT_PREVIEW=true only in local development or a Vercel Preview when you intentionally need to inspect draft content. Production ignores that switch. The feedback endpoint fails closed when Turnstile, distributed rate limiting or email delivery is missing.

Quality gates

Every code change is expected to pass the core local gate:

npm run lint
npm run typecheck
npm test
npm run validate:data
npm run validate:d1
npm run validate:manifests
npm run build
npm run test:e2e

CI additionally validates Source Manifests, prompt-injection fixtures, maintenance capacity, all Worker test suites and Worker deployment dry-runs. The complete executable matrix lives in .github/workflows/ci.yml.

Data changes must satisfy a stricter rule: an increase in row count is not progress unless source coverage, current-cycle coverage, field completeness and update success improve with it.

Repository map

.github/workflows/                 CI, source checks, backups and deployments
content/data/                      generated JSON compatibility snapshot
content/source-manifests/          allowlisted official-source manifests
docs/                              architecture, operations and data policies
infra/d1/                          Pipeline and Catalog D1 migrations
scripts/catalog/                   release building and performance benchmarks
scripts/ingestion/                 manifest, source and materialization tooling
scripts/quality/                   coverage, inventory and platform scorecards
src/app/[locale]/                  localized App Router pages
src/app/api/v1/                    versioned public API compatibility routes
src/components/                    design system and product features
src/i18n/                          locale registry and reviewed interface copy
src/lib/catalog/                   json / shadow / d1 repository implementations
src/lib/data/                      schemas, publication gates and formatting
tests/unit/                        domain, API and browser-storage tests
tests/e2e/                         multilingual critical-path tests
workers/ingestion/                 fetch, snapshot, parse and extraction pipeline
workers/entity-materializer/       candidate-to-entity field mapping
workers/publisher/                 validated publication candidates
workers/release-builder/           immutable release assembly and cutover
workers/catalog-api/               public D1 API Worker
workers/localization/              gated translation pipeline (disabled by default)

Release, rollback and recovery

• main is the production code branch; every pull request receives a Vercel Preview.
• A catalogue release is immutable and moves live only after schema, relationship, count, search and checksum validation.
• Catalog D1 retains the active release and two rollback releases; full history remains in private R2.
• A data rollback moves currentReleaseId back to a validated release without redeploying the application.
• An application rollback promotes the previous healthy Vercel deployment.
• Daily exports target a 24-hour recovery point; recovery restores into an isolated database before any production cutover.

See docs/platform-rollout.md, docs/backup-and-restore.md and docs/database-schema.md for operational detail.

Roadmap: depth before uncontrolled scale

flowchart LR
    N["Now<br/>266 public universities"] --> D["Data depth<br/>70% current-cycle coverage"]
    D --> T["Trust coverage<br/>266 manifests + reconciliations"]
    T --> C["D1 cutover<br/>3 releases / 72h shadow parity"]
    C --> F["500 universities<br/>after 2 healthy monthly cycles"]
    F --> K["1000+ universities<br/>quality- and cost-gated"]

Loading

Near-term work is measured by:

• raising the remaining 8 sparse universities to 3–5 verified international-student programs or a documented limited reconciliation;
• increasing current-cycle coverage from 20.50% to at least 70%;
• reaching 90% duration, 80% official application-route and 95% teaching-language coverage;
• expanding scholarship-connected institutions from 207 to at least 230;
• completing 266 Source Manifests and 266 catalogue reconciliations;
• completing three matching shadow releases over at least 72 hours before Production switches to D1;
• passing two full monthly update cycles before expansion to 500, then 1,000+ institutions.

Contributing data safely

Corrections and additions are welcome when they strengthen evidence quality.

1. Provide an official HTTPS URL from the university, government or scholarship provider.
2. Identify the academic year and intake for any deadline, fee or requirement.
3. Include the date on which the source was checked.
4. Leave unannounced values empty; do not infer them from an earlier year.
5. Never use a ranking, agency or aggregator as the sole evidence for an admissions fact.
6. Never commit passports, transcripts, health records, applicant emails or other personal application data.
7. Run lint, typecheck, tests and data validation before opening a pull request.

Use the data-correction issue form or follow the pull-request template. Raw collection artifacts and candidate packages belong in the private pipeline/quarantine flow, not in the public catalogue.

Security and privacy boundaries

• Collection is limited to registered official HTTPS domains and respects access controls, robots policies and per-domain throttling.
• The platform does not bypass authentication, CAPTCHAs, 403 responses, regional restrictions or private endpoints.
• Fetch validation rejects private-network addresses, unusual ports and unregistered cross-domain redirects.
• Applicants do not create accounts and the project does not collect application documents.
• Favorites and comparison state remain in the browser.
• Feedback is protected with origin checks, Turnstile, HMAC-based rate limiting and private email delivery.

Report a security problem privately to the maintainer rather than placing secrets or exploit details in a public issue.

Creator

Created and maintained by Henry Yang as a non-commercial public-interest information project for international students.

If you spot outdated information, use the website’s private correction form and include the official source. Please never send passports, transcripts, medical records or payment information.