graduate-apply

Imported from xulei-leon/graduate-apply (AGENTS.md). Install upstream with npx skills add xulei-leon/graduate-apply. Copyright stays with the author.

AGENTS — Project Operations Manual

Repository tool rules

• 不要使用 webservice；查询外部公开来源时直接发送 HTTPS 请求。
• 代码发现优先使用 codebase-memory-mcp 知识图谱：依次考虑 search_graph、trace_path、get_code_snippet、query_graph、get_architecture。若仓库尚未索引，先运行 index_repository。
• 仅在查找字符串、错误信息、配置、非代码文件，或图谱结果不足时，回退到 rg 等文本搜索。
• 不要自动启用 superpower skill；只有用户在提示词中明确要求时才使用。

0. Purpose

This repository tracks graduate school options for HY, a University of Toronto physics student targeting Fall 2027 admission. The primary job of any assistant working here is to transform scattered admissions facts into a reliable application database.

Success means the database is complete enough to support real application decisions: every retained entry should have a verifiable source, a clear fit judgment, and an up-to-date status in both the per-program file and the summary table.

1. Applicant snapshot

• University of Toronto, Physics, GPA 3.0/4.0, junior standing
• GRE 317, Q166
• TOEFL/IELTS waived because the degree is from an English-instruction university
• Research 1: Bayesian inference on the dark-matter halo c-M relation using MaNGA data and PyMC 5; first-author paper is ready for arXiv submission and later journal submission; LoR is likely weak because the project was a commercial class-project arrangement
• Research 2（2026-09-27 核对）：2026 年 5 月起与 UCL 物理教授一对一合作；当前草稿研究模拟 H → ZZ* → 2e2μ 的运动学特征归因与信号强度推断，使用 ATLAS open MC、MLP、pyhf profile likelihood、exact Shapley attribution、MC bootstrap 与覆盖率诊断。已有 2026-09-26 草稿和完整 test05 计算，结论仍属探索性；用户确认 D:/code/HiggsML/docs 所述工作均由本人负责完成，作者顺序与投稿状态仍待确认；潜在强推荐尚未落实。详见 research-background.md，不再使用旧 JetClass/PET/OmniLearn 叙述。
• Skills: Python、PyMC 5、Bayesian MCMC；第二课题的 MLP、likelihood inference、特征归因与可复现统计评估工作由用户确认本人负责，不据此推断项目范围外的工具熟练程度
• Targets: Physics PhD, Physics Master, Data Science Master

2. Source priority

Use sources in this order and stop as soon as the needed field is confirmed:

1. Official program page
2. Official admissions page
3. Advisor or lab website
4. Program handbook or FAQ
5. Graduate school admissions page
6. Third-party aggregators only as backup context, never as the primary source

If two sources disagree, keep both in the notes and prefer the newer official source.

If the needed field cannot be confirmed from sources in this order, leave it unresolved rather than inferring.

3. Required data by program type

3.1 Master programs

For every Master entry, capture these fields in the relevant folder README and in a per-program file:

Required file pattern for each Master program:

• Program header with university, degree, track, and region
• Fact table covering the fields above
• Source list with visible URLs
• Last-verified date
• Short fit note that explains the match level in one sentence

3.2 PhD programs and advisors

For every PhD program, capture the program-level fields above plus advisor records.

Program-level fields:

• Program name, department, degree type
• GPA / GRE / TOEFL requirements
• Cohort size
• Deadline, including fellowship-priority deadlines
• Funding model, including whether self-funded study is accepted
• Whether direct entry from bachelor's is accepted (direct-entry PhD)
• Official URL

Required file pattern for each PhD advisor file:

• Program header with degree, department, and region
• Program-level fact table
• Advisor profile with research fit and current status
• Source list with visible URLs
• Last-verified date
• Short outreach note if contact is needed or has happened

Advisor-level fields:

• Name, title, department
• Research focus, narrowed to the subfield level
• Current projects or grants
• Representative papers from the last 3 years
• Whether the lab is recruiting
• Whether outreach is needed
• Contact status: not contacted, emailed, replied, positive, negative

4. Match logic

Use this scale consistently across programs and advisors:

If the evaluation depends on a single missing fact, mark it as unknown rather than upgrading the fit.

When in doubt, choose the lower match level until the missing fact is verified.

5. Country and GPA interpretation

• United States: GPA 3.0 is below the common PhD/master applicant average, so research and writing quality matter more than ranking alone
• Canada: GPA 3.0 is often within range for many Master programs, especially at U of T peers or closely related schools
• UK: interpret GPA through degree classification; U of T 3.0 is roughly comparable to a 2:1 range for many institutions
• Switzerland: course fit matters heavily; use strict prerequisite checking
• Singapore, Hong Kong, Australia: U of T recognition is usually helpful, but funding and prerequisite fit still matter

Do not use Safe/Match/Reach as a guarantee. These labels are only planning aids.

6. Output and file rules

• Create one markdown file per school or advisor before updating the folder summary table
• Use the naming convention already defined in each folder
• Keep source URLs visible in every file
• Add a last-verified date when a source is checked
• Update contact status as soon as it changes
• High-priority or deadline-nearing items should be visually marked in the summary table
• Preserve original-source quotations, official program and policy names, and source-language terminology in English. Write synthesized reports, comparisons, recommendations, and analytical prose in Chinese.
• Do not mark an entry complete until the per-program file and the index row both exist and agree on the key facts
• Keep summary counts, region totals, and target totals synchronized with the underlying files

7. Search prompts

Use these query patterns when starting a new search:

• Physics PhD: site:edu "Physics PhD" admissions requirements
• Physics Master: site:edu "Master of Science" Physics
• Canada: site:ca "MSc" Physics admission requirements
• UK: site:ac.uk "MSc" Physics entry requirements
• Switzerland: site:ch "Master" Physics ETH or site:ch "Master" Physics EPFL
• Singapore: site:edu.sg "Master" Physics or site:edu.sg "PhD" Physics
• Hong Kong: site:edu.hk "MSc" Physics
• Australia: site:edu.au "Master" Physics
• Data Science: site:edu "Data Science" Master admission requirements
• Advisor search: dark matter Bayesian professor physics, galaxy dynamics MCMC professor, computational astrophysics faculty

8. Quality bar

Every entry should make the next model's job easier by answering four questions without interpretation: can the applicant apply, what is required, how strong is the fit, and what source supports the claim. If a field cannot be verified from a reliable source, leave it explicitly unresolved.

9. Search workflow and targets

9.1 School ranking filter

Applies only to Master programs. PhD programs have no ranking restriction.

Canada exception: Do not apply the QS or US News ranking filter to Canadian universities. Canadian Physics Master and Data Science Master programs should be screened by academic eligibility, research fit, supervisor availability, funding, and applicant strategy instead of institutional rank.

For countries other than Canada, use the most recent published rankings. If a non-Canadian school does not meet this bar for Master programs, do not add it.

9.2 Quantitative targets per category

Treat these as minimum coverage goals, not hard ceilings. If a category still lacks regional balance, a strong match, or enough verified entries, continue searching until the gap is closed or explicitly document why it cannot be.

Distribute targets across the agreed regions proportionally. For example, Physics Master should cover all 7 regions, with more weight on the US and UK.

9.3 Per-category instructions

Physics PhD (physics-phd/)

• Search for programs in the US and Singapore (no ranking filter).
• Only consider programs that accept direct entry from bachelor's (direct-entry PhD).
• For every program, identify 1–3 potential advisors whose research overlaps with computational physics broadly. Priority areas: computational astrophysics (dark matter, galaxy dynamics, cosmology) plus computational methods (ML for physics, scientific computing, numerical simulations, data science).
• Create one file per advisor (not per program). If multiple advisors at the same university fit, create separate files for each.
• Not every university needs an advisor file — only create files for advisors whose research is a genuine match (high or medium).
• Record the program-level info in the advisor file's header section. The program index table tracks which universities have one or more matched advisors.

Physics Master (physics-master/)

• Search for physics or astronomy master programs (MSc, MPhys, MASt, etc.).
• Favour programs with a computational, data-analysis, or quantitative-skills component.
• Create one .md file per program.
• If a university offers multiple relevant tracks, you may combine them into one file or split as needed.

Data Science Master (ds-master/)

• Search for DS / Data Analytics / Applied Statistics master programs that accept STEM backgrounds.
• Prioritise programs that explicitly welcome non-CS majors or note that a physics background is sufficient.
• Create one .md file per program.

9.4 Step 1 — Search and capture

For each target program or advisor:

1. Open the official program / admissions / advisor page.
2. Extract all fields required by section 3.
3. Save to a single .md file in the correct folder.
4. Update the folder README.md index table with a summary row.
5. Move to the next target.

Before moving on, verify that the new file, the summary row, and any folder-level counts all match.

9.5 Step 2 — Summarise

After all targets in a category have been captured:

1. Ensure the folder README.md index table is complete and up to date.
2. For each entry, add or verify:
   • Match level (High / Medium / Low)
   • Difficulty (Safe / Match / Reach), using the definitions in section 5 and section 10 of the older guide (see country-differences.md for context)
   • Deadline — highlight any approaching deadlines
3. Sort the table by priority: Reach items first (they need the most preparation), then Match, then Safe.

Before closing a category, check that:

• All target rows have live source URLs
• Last-verified dates are present
• Region totals add up
• Match and difficulty labels are consistent with the evidence
• No entry relies on a guessed or stale requirement

9.6 Diversity requirement

Spread targets across regions and institution tiers. Do not concentrate all 30 Physics Master picks on US schools alone. Ensure at least 2–3 entries per region (Canada, UK, Switzerland, Singapore, Hong Kong, Australia) are present in the Master categories.