How this site is made

This site is a small static one, built with Astro and served from Cloudflare Pages. There is no database and nothing runs when you visit: every page you see was written to a file before it was published. The display face is Fraunces, self-hosted rather than loaded from a font service, so reading a page here does not report back to anyone. The accent is 清华紫, Tsinghua purple, which is where I studied.

How AI is involved

Most of the code behind this site was written by Claude, working in Claude Code, with me directing it: I decide what the site should do, argue with the proposals, and correct the results. That is the same way I now use AI in research, and it seemed dishonest to use it that way and not say so on the site it produced.

The writing is a different division of labour, and it varies from piece to piece. The ideas and the judgements are mine. Some pieces I draft and Claude helps revise; some Claude drafts from my notes and outline and I rewrite; every factual claim, number and citation is mine, and I check them. Where the method matters to reading a particular piece, that piece says so in a sentence of its own rather than wearing a badge. I considered a tiered label on every page, marking how much was mine and how much was drafted, and decided against it: almost nothing is cleanly one or the other, such a label stops being true the first time I edit, and a ladder from "human" down to "AI" implies a ranking I do not believe in. AI here is an instrument, the way a microscope is an instrument, and what matters is whether the result is right.

Where the information comes from

The publication list is generated from structured records of my own papers, so the titles, authors and abstracts are as published rather than retyped. Whether something is open access, and whether an author copy exists in a repository, comes from OpenAlex and is refreshed rather than asserted, which is why a paper may be marked as having an accessible version even when the journal itself is paywalled. Links on people's names go to their ORCID records; where I have no reliable profile for someone, their name is not a link, because a wrong link points at the wrong person.

Full texts are not reproduced here. I built a reader for them and then took it out: the text extracted from the published PDFs carried enough damage, with journal furniture spliced into the prose, that it misrepresented the papers. The publisher's version is free in most cases, is correctly typeset, and is the citable one, so every entry sends you there instead.

What this site does not do

There is no analytics, no tracking and no advertising; nothing here counts you or follows you. The only thing stored in your browser is your choice of light or dark theme, kept on your own device. There is no email address on the site, deliberately, and the build strips any that appear inside quoted material; you can reach me through the profiles linked on the front page.